50 Small-Business Processes Ranked for AI Agents: Where to Start

Decentralised News Research · Small-business workflows

What Matters

The best first AI workflow has a clear output, a visible baseline and a mistake the business can contain. DN’s 50-process ranking helps small businesses shortlist work for a controlled pilot. It does not establish measured returns. Start with the process, define what the system may do, and measure accepted outcomes, remaining human work and costs before expanding its authority.

Choose the process before the platform

A business does not automate “sales” or “finance” in one purchase. It changes particular activities: preparing a meeting brief, classifying an invoice, drafting a reply or requesting a payment. These activities require different information, permissions and standards of correctness. A product that handles one well may be unsuitable for the next.

Begin with a recurring bottleneck whose inputs, owner and finished output you can describe. A useful candidate might produce draft descriptions from an approved product catalogue. “Handle our marketing” leaves the acceptance standard, customer exposure and spending authority undefined.

Anthropic’s engineering guidance distinguishes workflows with predefined paths from agents that dynamically select steps and tools. It advises starting with the simplest sufficient approach, because additional autonomy can add cost and latency. These are architecture principles, not evidence of small-business payback. A form, rule or existing automation may already solve the bounded task. [1]

Once the workflow is defined, DN’s small-business automation tools guide provides product context. A tool list cannot decide whether your particular process is ready.

Separate assistance from authority

Describe the intended operating mode before evaluating a demonstration. Four distinctions make the scope reviewable:

• Assist: retrieve, extract or summarise information for a person who remains responsible for the next action.
• Draft: prepare a proposed message, record or document without sending, publishing or committing it.
• Recommend: propose a decision with supporting evidence, uncertainty and an explicit approval point.
• Execute: change an external system within a specified permission boundary, such as creating an approved draft record.

Approval is a separate control, not a tone of voice in the output. A system that writes “please review” while holding unrestricted payment credentials still has consequential authority. Define who can approve, what they inspect and which downstream system actually enforces the decision.

Prefer narrow permissions that match the task. OWASP’s excessive-agency guidance recommends minimum functions and permissions, approval for high-impact actions and authorization enforced downstream. It specifically distinguishes reading messages from having the ability to send or delete them. [3]

Read the 50-process ranking as a starting order

The ranking is DN’s editorial screening framework. It considers process characteristics to help a small business build a shortlist; it is not a league table of measured savings, provider quality or deployment success. Its assumptions should be examined against your records, systems and staff capacity.

DN SME Opportunity Finder v1.0 covers 50 selected processes and zero measured deployments or ROI results. Its default business uses cloud applications, stable processes, modest repeat volumes, authorized data and human review. Four editorial ratings run from one to five: value potential (V), clarity of acceptance criteria (C), integration effort (E, where five is hardest) and consequence of error (R, where five is most serious).

Priority score = 35 × (V − 1) ÷ 4 + 25 × (C − 1) ÷ 4 + 20 × (5 − E) ÷ 4 + 20 × (5 − R) ÷ 4.

The 0–100 score is starting priority, not a performance percentage. Reproducible arithmetic does not validate the underlying judgments. Time to ROI remains Unmeasured.

Hard gates override the score. Payments, refunds, bank-detail changes, account deletion, recruitment ranking, payroll, tax submission and access grants belong in Protected controlled pilots, not first pilots. Other processes qualify as First pilot only with R ≤ 2, E ≤ 2 and C ≥ 4; remaining processes are Supervised. These classifications do not grant execution authority.

A high-ranked process can still be unsuitable when the underlying data is unreliable or no one can review the output. A lower-ranked process may deserve a carefully bounded pilot when it addresses an expensive, well-understood bottleneck. Treat unavailable evidence as a question to resolve, rather than a favourable default.

Version 1.0 rating definitions and ordering

V — Value potential

1: incidental convenience; 2: occasional administrative help; 3: useful recurring support; 4: meaningful recurring operating workload; 5: potentially material workload or control benefit. Actual volume and realised value require local measurement.

C — Clarity of acceptance criteria

1: subjective outcome without a simple check; 2: ambiguous criteria; 3: mixed objective and judgmental criteria; 4: mostly explicit checks; 5: clear schema, approved source or deterministic comparison.

E — Integration effort

1: one controlled input and draft output; 2: a maintained source plus an existing application; 3: coordinated records or multiple read integrations; 4: multiple changing systems and exceptions; 5: consequential cross-system authority or substantial custom work. These ratings are not estimates of hours or cost.

R — Consequence of error

1: low-stakes reversible internal categorisation; 2: contained rework with review before effects; 3: potential customer, financial-record or policy impact; 4: significant financial, service or operational consequences; 5: severe or difficult-to-reverse money, rights, access, deletion or binding-commitment effects. Ratings describe the process with its prescribed controls, not the probability of harm.

Gating comes before score: non-protected records precede Protected records. Within those groups, sort by score descending, V descending, R ascending, E ascending, then stable process ID ascending. Display ranks are unique; equal scores remain tied in evidence.

The weights and thresholds are declared editorial choices. Small score gaps do not establish meaningful differences, and the list is not a representative survey of businesses.

Use six gates before starting a pilot

1. Define an accepted outcome

Write a short acceptance checklist that another person can apply. For invoice extraction, that could include matching the supplier, amount and reference to the source document. Retain rejected outputs and correction time. A fluent response or completed tool call is insufficient when the required business result is wrong.

2. Establish the data boundary

Identify the necessary sources, who may access them and what information may leave the existing system. Separate approved business instructions from customer messages and retrieved documents. An external document should not be able to grant new permissions. Resolve sensitive-data handling and retention questions before feeding real records into a pilot.

3. Limit what can change

List permitted actions and prohibit everything outside that list through account permissions and application controls. Specify relevant amounts, recipients, record types and environments. Drafting a refund explanation should not imply permission to issue a refund. Prompt instructions can describe a boundary; they do not establish the access control that enforces it.

4. Make approval meaningful

Give the reviewer the source, proposed change and reason for escalation. Define who covers an absent approver and what happens when approval expires. Avoid a queue in which every item receives the same automatic click. Count review time as part of the workflow, and escalate when the evidence does not support a decision.

5. Demonstrate recovery

Decide how to stop new actions, reconcile work already attempted and restore the previous process. Preserve operation identifiers and final-state evidence. A timeout may leave a completed external action: DN’s transaction failure analysis explains why an error message alone cannot justify retrying. Some actions need compensation rather than reversal.

6. Assign an accountable owner

Name the person who reviews exceptions, costs and changes to the system. Set review dates and stopping conditions before launch. NIST’s voluntary AI Risk Management Framework treats risk management as an ongoing activity spanning design, use and evaluation. It does not certify this ranking or substitute for your own operational ownership. [2]

Build your shortlist

Use the finder to narrow the process list, then apply the six gates to the candidates that fit your business. Choose a small enough pilot that a named person can inspect its outputs and restore the previous workflow. The finder organises editorial assessments; it cannot inspect your records, verify permissions or establish financial returns.

DN SME Opportunity Finder

Combine a function and pilot class to narrow the table. Reset restores all 50 processes.

Business function

All functions Customer support Sales Finance Marketing People and IT Operations

Pilot class

All classes First pilot Supervised Protected

Reset to all processes

Filtering is unavailable in this browser. All 50 processes remain visible with their function and pilot class.

All 50 processes50 processes · All functions · All classes19 processes · All functions · First pilot19 processes · All functions · Supervised12 processes · All functions · Protected8 processes · Customer support · All classes5 processes · Customer support · First pilot0 processes · Customer support · Supervised3 processes · Customer support · Protected8 processes · Sales · All classes3 processes · Sales · First pilot3 processes · Sales · Supervised2 processes · Sales · Protected9 processes · Finance · All classes0 processes · Finance · First pilot6 processes · Finance · Supervised3 processes · Finance · Protected8 processes · Marketing · All classes6 processes · Marketing · First pilot1 process · Marketing · Supervised1 process · Marketing · Protected7 processes · People and IT · All classes2 processes · People and IT · First pilot3 processes · People and IT · Supervised2 processes · People and IT · Protected10 processes · Operations · All classes3 processes · Operations · First pilot6 processes · Operations · Supervised1 process · Operations · Protected

No processes match these filters. Choose another function or pilot class, or reset to all processes. An empty shortlist is not a readiness score for your business.

DN SME Opportunity Finder v1.0 · all 50 editorial records

Rank	Process, function and approach	Score / 100 V · C · E · R	Pilot class	Control, validation and rationale
1	Detect missing fields in incoming formsOperations · Rules first P43	91.254 · 5 · 1 · 1	First pilot	Control: Keep the original submission and a clear correction route.Validate: Compare flags with the form schema and known valid examples. Why this rating? A deterministic completeness check is highly observable and may not need an LLM.
2	Draft answers from an approved FAQCustomer support · LLM assist P01	90.005 · 5 · 2 · 2	First pilot	Control: Human reviews the answer before sending.Validate: Compare every claim with the approved source; measure edits and reopenings. Why this rating? Frequent bounded questions offer visible value when a maintained knowledge base exists.
3	Turn approved meeting notes into action listsSales · LLM assist P09	86.254 · 5 · 1 · 2	First pilot	Control: Meeting owner confirms tasks before assignment.Validate: Compare owners, deadlines and commitments with the approved notes. Why this rating? A narrow output format and readable source make verification simple.
4	Repurpose approved content into draft postsMarketing · LLM assist P26	86.254 · 5 · 1 · 2	First pilot	Control: Editor approves publication and any new claim.Validate: Check every factual claim and preserve the approved message. Why this rating? Known source material and a narrow transformation make this a practical first pilot.
5	Extract invoice fields for reviewFinance · Document extraction P17	85.005 · 5 · 2 · 3	Supervised	Control: The reviewer approves extraction; no ledger posting or payment.Validate: Check supplier, invoice ID, amount and currency against the original. Why this rating? High repetition and structured fields create value, while monetary errors remain material.
6	Classify and route support ticketsCustomer support · Rules plus LLM P02	81.254 · 5 · 2 · 2	First pilot	Control: Escalation queue remains visible to a person.Validate: Compare proposed queues with an adjudicated sample, including urgent exceptions. Why this rating? A stable taxonomy makes routing errors observable and reversible.
7	Draft order-status repliesCustomer support · Rules plus LLM P04	81.254 · 5 · 2 · 2	First pilot	Control: Use authoritative order data; never invent delivery dates.Validate: Check the order ID, current status and recipient before every send. Why this rating? Structured status lookup is bounded, but stale or mismatched records can mislead customers.
8	Propose CRM field updates from correspondenceSales · LLM assist P10	81.254 · 5 · 2 · 2	First pilot	Control: Approve a change list; do not silently overwrite CRM data.Validate: Check each proposed field and provenance against the source. Why this rating? Structured fields enable quick review when the existing CRM schema is consistent.
9	Draft product descriptions from approved attributesMarketing · LLM assist P27	81.254 · 5 · 2 · 2	First pilot	Control: Human approves; forbid invented specifications or certifications.Validate: Compare specifications and claims with the product catalogue. Why this rating? Structured product facts support repeatable drafting when the catalogue is current.
10	Summarise daily internal operations reportsOperations · Rules plus LLM P35	81.254 · 5 · 2 · 2	First pilot	Control: Read-only summary with links to original records.Validate: Reconcile figures to the reports; inspect missing or stale inputs. Why this rating? A regular reporting format makes time saved and factual accuracy observable.
11	Summarise long complaint threadsCustomer support · LLM assist P03	80.004 · 4 · 1 · 2	First pilot	Control: Keep the original thread alongside the summary.Validate: Check omitted commitments, disputed facts and dates against the thread. Why this rating? Read-only summaries reduce reading effort without deciding the complaint.
12	Draft standard procedures from approved notesOperations · LLM assist P34	80.004 · 4 · 1 · 2	First pilot	Control: Publish only after the owner verifies the procedure.Validate: Have the process owner walk through the proposed steps. Why this rating? Existing notes and an identifiable owner allow a low-integration pilot.
13	Draft appointment-confirmation repliesCustomer support · Rules first P05	77.503 · 5 · 1 · 2	First pilot	Control: A person resolves ambiguity; deterministic reminders may need no AI.Validate: Check time zone, service and availability against the booking record. Why this rating? Highly structured messages are easy to validate with modest potential value.
14	Draft newsletter summaries from owned contentMarketing · LLM assist P28	77.503 · 5 · 1 · 2	First pilot	Control: Editor approves the final send.Validate: Verify links, claims and consistency with source articles. Why this rating? Owned source content lowers research effort and makes checking straightforward.
15	Draft image alt text for a human editorMarketing · LLM assist P31	77.503 · 5 · 1 · 2	First pilot	Control: A person checks accuracy, context and accessibility.Validate: Compare descriptions with the actual image and page purpose. Why this rating? The output is bounded, but context still determines a useful description.
16	Draft onboarding checklists from an approved role templatePeople and IT · Rules plus LLM P44	77.503 · 5 · 1 · 2	First pilot	Control: Manager approves; do not grant access automatically.Validate: Check required steps and accountable owners against the approved template. Why this rating? A bounded checklist has low integration effort and visible completion criteria.
17	Draft overdue-invoice remindersFinance · Rules plus LLM P20	76.254 · 5 · 2 · 3	Supervised	Control: Approve sending and suppress reminders after payment or dispute.Validate: Check payment status, amount, recipient and disputed balances. Why this rating? Useful recurring work, provided current ledger state prevents inappropriate reminders.
18	Flag low-stock items for reviewOperations · Rules first P36	76.254 · 5 · 2 · 3	Supervised	Control: A person decides replenishment; fixed thresholds may be sufficient.Validate: Compare stock, lead time and open orders with the inventory system. Why this rating? Simple stock alerts are valuable and often need ordinary automation before an agent.
19	Extract dates from routine contracts for a registerOperations · Document extraction P41	76.254 · 5 · 2 · 3	Supervised	Control: Responsible owner verifies entries; this is not legal interpretation.Validate: Verify each date and triggering clause against the signed document. Why this rating? Bounded extraction is useful, but missing a renewal date can have material consequences.
20	Draft follow-up emails after a meetingSales · LLM assist P12	75.004 · 4 · 1 · 3	Supervised	Control: A person approves each external message.Validate: Check commitments, tone, recipients and factual claims. Why this rating? Drafting is easy to start; commercial promises and relationship context need review.
21	Flag possible duplicate invoicesFinance · Rules plus LLM P18	73.755 · 4 · 3 · 3	Supervised	Control: Treat flags as candidates, not instructions to withhold payment.Validate: Review flagged and missed duplicates against invoice records. Why this rating? Cross-record matching can reduce rework but needs reliable identifiers and careful exceptions.
22	Draft purchase requisitionsOperations · Rules plus LLM P38	71.254 · 5 · 3 · 3	Supervised	Control: Approval is required before any purchase order is issued.Validate: Check requested items, budget coding and quantities. Why this rating? Structured requests can save effort once procurement data is connected.
23	Cluster search terms by intentMarketing · LLM assist P29	71.253 · 4 · 2 · 1	First pilot	Control: Use clusters as planning suggestions, not search-volume measurements.Validate: Review clusters against actual pages and customer questions. Why this rating? Low-consequence internal categorisation is reversible and easy to inspect.
24	Summarise win-loss interview notesSales · LLM assist P15	71.253 · 4 · 1 · 2	First pilot	Control: Remove unnecessary identifiers and validate interpretation.Validate: Trace themes to the interviews; retain contradictory feedback. Why this rating? Useful qualitative synthesis is easy to review but is not causal sales evidence.
25	Triage inbound sales enquiriesSales · Rules plus LLM P11	70.004 · 4 · 2 · 3	Supervised	Control: People review low-confidence and rejected enquiries.Validate: Check routing and missed enquiries against a written qualification rubric. Why this rating? Routing can save time, but a bad qualification decision can lose an opportunity.
26	Propose expense categoriesFinance · Rules plus LLM P19	70.004 · 4 · 2 · 3	Supervised	Control: An authorised reviewer approves coding before posting.Validate: Compare proposed categories with the business's accounting policy. Why this rating? Repeatable categories help, but ambiguous expenses require business context.
27	Normalise supplier quotes into a comparisonOperations · LLM assist P37	70.004 · 4 · 2 · 3	Supervised	Control: Human validates the comparison and makes the purchasing decision.Validate: Check unit price, minimum order, delivery terms and exclusions. Why this rating? Formatting differences are a useful AI task, but omitted terms can change the decision.
28	Retrieve approved internal policy passagesPeople and IT · LLM assist P45	70.004 · 4 · 2 · 3	Supervised	Control: Escalate ambiguous policy questions and restrict retrieval by permissions.Validate: Check answer citations, document version and access entitlements. Why this rating? Retrieval is valuable, but wrong policy advice or excess document access matters.
29	Draft internal IT troubleshooting stepsPeople and IT · LLM assist P46	70.004 · 4 · 2 · 3	Supervised	Control: No privileged commands; staff approve actions and escalation.Validate: Match suggestions to the approved knowledge base and supported systems. Why this rating? A bounded support library supports assistance while avoiding unsupervised system changes.
30	Summarise authorised public customer reviewsMarketing · LLM assist P30	66.253 · 4 · 2 · 2	First pilot	Control: Do not infer private personal attributes or invent quotations.Validate: Link themes to source reviews and check sampling bias. Why this rating? A useful discovery task with manageable consequences when sources are retained.
31	Draft job descriptions from an approved role briefPeople and IT · LLM assist P47	66.253 · 4 · 1 · 3	Supervised	Control: Human approves the job advert and selection criteria.Validate: Manager checks accuracy, essential requirements and wording. Why this rating? Drafting can help, but employment context and inaccurate requirements demand review.
32	Propose bank-reconciliation matchesFinance · Rules first P22	63.755 · 4 · 4 · 4	Supervised	Control: An accountant approves matches; no autonomous journal posting.Validate: Check amount, date, reference and unmatched items in the source ledger. Why this rating? Potential value is high, but exceptions and multi-system access increase effort and risk.
33	Summarise non-sensitive training feedbackPeople and IT · LLM assist P50	62.502 · 4 · 1 · 2	First pilot	Control: Aggregate results; do not infer individual performance or protected traits.Validate: Trace themes to consented feedback and check omitted criticism. Why this rating? Low-stakes internal synthesis is simple to test, with limited direct financial value.
34	Prepare a cash-position summaryFinance · Rules plus LLM P21	60.004 · 4 · 3 · 4	Supervised	Control: Read-only reporting; a finance owner validates every total.Validate: Reconcile totals to approved source reports and label missing accounts. Why this rating? Visibility is valuable, but incomplete balances can distort financial decisions.
35	Prepare public-source account briefsSales · Bounded research agent P13	60.003 · 3 · 2 · 2	Supervised	Control: Restrict research to authorised public sources and a bounded task.Validate: Require dated source links and verify important facts before use. Why this rating? Discovery involves branching, but stale or misattributed facts limit reliability.
36	Generate advertising variants from approved copyMarketing · LLM assist P32	55.003 · 3 · 2 · 3	Supervised	Control: Human approves copy, targeting and spend.Validate: Check factual claims, audience context and brand rules before any experiment. Why this rating? Generating variants is simple; public claims and commercial impact need controls.
37	Suggest reorder quantitiesOperations · Rules first P39	48.754 · 3 · 4 · 4	Supervised	Control: Buyer approves quantities and commitments.Validate: Back-check suggestions against demand, lead times and stockouts. Why this rating? Forecast uncertainty and inventory costs require more validation than basic stock alerts.
38	Propose dispatch or scheduling changesOperations · Rules plus LLM P40	48.754 · 3 · 4 · 4	Supervised	Control: Dispatcher approves all schedule changes.Validate: Check capacity, travel constraints and affected customer commitments. Why this rating? Multiple live constraints make reliable integration and exception handling substantial work.
39	Draft quotes from an approved price listSales · Rules plus LLM P14	66.254 · 5 · 3 · 4	Protected	Control: Human signs off every quote; calculations use deterministic rules.Validate: Check SKU, tax treatment, validity, price and terms with the quote owner. Why this rating? A quote can create costly commitments even when its wording looks routine.
40	Release payroll or supplier paymentsFinance · Rules first P24	65.005 · 5 · 4 · 5	Protected	Control: Named authorisers approve through existing payment controls.Validate: Validate payee, amount, duplicate protection and reconciliation evidence. Why this rating? Clear arithmetic does not remove the high consequence of an incorrect transfer.
41	Prioritise sensitive or urgent complaintsCustomer support · Rules plus LLM P06	62.505 · 3 · 3 · 4	Protected	Control: A human owns urgency decisions and the exception queue.Validate: Review missed urgent cases and false escalations with the responsible owner. Why this rating? The cost of a missed urgent case outweighs convenience; do not make this the first autonomous pilot.
42	Approve refunds and compensationCustomer support · Rules first P07	60.004 · 4 · 3 · 4	Protected	Control: Human authorises money movement under a documented policy.Validate: Reconcile amount, policy, identity and previous refunds. Why this rating? Financial consequences and duplicate refunds require transaction controls.
43	Execute supplier bank-detail changesFinance · Rules first P23	50.004 · 4 · 4 · 5	Protected	Control: Separate verification and approval; do not trust emailed instructions alone.Validate: Independently verify the change through the established supplier process. Why this rating? A convincing message can redirect funds, making autonomous execution unsuitable as a first pilot.
44	Submit tax returns or statutory filingsFinance · Rules first P25	50.004 · 4 · 4 · 5	Protected	Control: Qualified review and explicit approval precede any filing.Validate: Check source totals, submission status and applicable requirements with the responsible professional. Why this rating? Submission creates consequences that cannot be judged from fluent text alone.
45	Approve access requests or privilege changesPeople and IT · Rules first P49	50.004 · 4 · 4 · 5	Protected	Control: Use least privilege and a named access approver.Validate: Check requester identity, role, resource and approval record. Why this rating? An incorrect grant can expose data or systems beyond the immediate task.
46	Change advertising budgets autonomouslyMarketing · Rules first P33	43.754 · 3 · 4 · 5	Protected	Control: Authorised people set and approve spending limits and material changes.Validate: Check spend caps, attribution quality and stop conditions against actual results. Why this rating? Noisy attribution and money movement make this a poor first autonomous workflow.
47	Execute customer account deletionCustomer support · Rules first P08	43.752 · 5 · 3 · 5	Protected	Control: Use the existing verified deletion process with human authorisation.Validate: Verify authority, required retention checks and an auditable deletion result. Why this rating? Deletion is hard to reverse and may affect records beyond the visible account.
48	Rank job candidates or reject applicantsPeople and IT · Rules first P48	37.504 · 2 · 4 · 5	Protected	Control: People remain responsible for consequential employment decisions.Validate: Use the organisation's governed hiring assessment and review process. Why this rating? Individual opportunity and fairness make this unsuitable for a first autonomous pilot.
49	Send autonomous outbound sales sequencesSales · Rules first P16	33.753 · 2 · 4 · 4	Protected	Control: Keep a named owner, approved audience and stop control.Validate: Review recipient eligibility, suppression lists, volume and complaint signals. Why this rating? Unsupervised messaging can damage relationships and amplify selection errors.
50	Auto-sign contracts or accept supplier termsOperations · Rules first P42	32.504 · 2 · 5 · 5	Protected	Control: An authorised human makes and records the commitment.Validate: Use the existing legal and commercial review process. Why this rating? Binding commitments require authority and contextual judgment beyond a first pilot.

Time to ROI is Unmeasured for every record. Scores are editorial priorities, not percentages of readiness or measured performance. Filtering does not recalculate or approve the ratings.

A practical 30-day pilot

This schedule is a planning example. Integration or approval work may require longer; the calendar is not evidence that a process is ready.

1. Days 1–5: document the baseline. Choose one process, an owner and the acceptance checklist. Record normal task mix, accepted outputs, human minutes, exceptions and direct costs. Identify the existing method that will remain available if the pilot stops.
2. Days 6–10: run without committing changes. Use permitted historical or synthetic cases to produce draft outputs. Include awkward cases, missing information and conflicting instructions. Confirm that the reviewer can identify errors and that permissions prevent out-of-scope actions.
3. Days 11–20: run a bounded comparison. Use comparable work in the baseline and pilot, with the same acceptance rules. Keep any necessary approval gates. Record every attempted case, review, correction and exception; excluding failed attempts would make the economics misleading.
4. Days 21–30: reconcile and decide. Check outputs against destination records, total costs and remaining work. Exercise the stop-and-recovery procedure. Expand, revise or discontinue the pilot based on evidence, with changes and unresolved questions recorded.

Stop when an agreed material condition occurs, such as an unauthorized action, unexplained record mismatch or inability to determine final state. Pause affected work, preserve evidence, reconcile pending actions and return to the baseline process where appropriate. Fixing a prompt is not sufficient evidence that the underlying control failure is resolved.

Measure accepted work and realised value

Compare equivalent workloads. Easy cases handled by automation and difficult cases left to people do not form a fair before-and-after comparison. Record the mix and any changes in volume, staffing or business demand. Distinguish a useful operational signal from a causal claim the comparison cannot support.

At minimum, retain attempted and accepted outcomes, human handling time, correction effort, escalations and actual recurring costs. Measure end-to-end completion where waiting matters. Ten minutes of faster drafting may offer little customer benefit when approval still takes three days.

Time to ROI remains unknown until costs, benefits and their timing are established. Freed employee capacity may improve service without reducing spending. Record separately whether the gain avoids overtime or contractor costs, produces attributable contribution margin, or remains unused capacity. DN’s industry ROI framework and calculator explain that conversion in more detail.

A cost worksheet with an explicitly fictional example

Keep one-time setup separate from recurring operation. Include integration, staff training and launch work in setup; include usage, maintenance and the remaining operating burden in the monthly comparison. Avoid counting the same review labour or benefit twice.

Net capacity hours per month = monthly volume × (baseline human minutes − human minutes with automation) ÷ 60.

Monthly net benefit = realised labour benefit + attributable incremental contribution margin − additional error losses − incremental recurring costs.

Simple payback in months = setup cost ÷ positive monthly net benefit. With zero or negative monthly net benefit, the model does not recover setup cost. This simplified calculation excludes ramp-up, discounting and taxes.

Suppose 600 monthly tasks fall from 12 to eight human minutes each, including review and correction. That frees 40 hours. At 30 currency units per hour, the capacity value is 1,200. If 50% becomes avoided expenditure, the realised labour benefit is 600.

Add 200 of incremental contribution margin, then subtract 50 of additional error loss and 350 of recurring costs. The assumed monthly net benefit is 400. A setup cost of 2,400 then gives six months of steady-state simple payback. Every input is fictional; this is not an industry benchmark.

With no realised labour saving, the same example loses 200 per month despite freeing time. If automation increases human time, charge that additional labour at its full cost instead of discounting it by a savings-realisation percentage. Keep uncertain benefits visible and test the case without them.

When to configure, buy or build

First check whether a tool you already use can support the bounded process with appropriate permissions and records. A new subscription is justified by the workflow it improves, not by duplicating a feature that staff already have.

Buying may suit a standard process when integrations, support and export are adequate. Building may make sense when the workflow or control requirements are distinctive and someone can maintain it. Neither route removes the cost of review, exception handling, access management or future changes.

Ask a supplier to demonstrate your acceptance cases, failure handling and authorization boundary. Review data handling, record export, usage limits, recurring charges and exit arrangements. Test whether a human can continue the work when the service is unavailable. Compare total operating obligations, not just the advertised subscription.

Keep commercial selection separate from readiness. A well-ranked process does not establish that a particular product, contract or integration is suitable. Begin with a bounded evaluation that can answer those questions.

Frequently asked questions

Are these measured rankings of business ROI?

No. They are editorial process assessments for shortlisting pilots. DN has not measured comparable deployment returns for all 50 processes. Financial payback requires evidence from the actual workflow and its full cost boundary.

Does every process need an autonomous agent?

No. A fixed rule, template or existing workflow may be sufficient. Use dynamic planning when the task needs it and its costs and failure modes can be managed.

What should a small business automate first?

Choose a recurring, bounded process with accessible inputs, a clear acceptance rule, a responsible reviewer and a recoverable failure. Resolve missing permissions or evidence before relying on its ranking.

Can a human approval step make any workflow safe?

No. The reviewer needs relevant evidence, time and genuine authority. The system must enforce the approval boundary, and the business still needs controls for mistakes, unavailable reviewers and unresolved actions.

Are hours saved the same as cash saved?

No. Time becomes a financial benefit only through a supported change such as avoided expenditure or attributable incremental margin. Spare capacity can still be useful, but should be reported separately.

Why is time to ROI not a fixed number for each process?

Implementation effort, task volume, acceptance rates, costs and benefit realisation differ by business. A process category alone cannot determine a reliable payback period.

What happens when an action times out?

Keep its outcome unresolved until destination records establish what happened. Preserve the original identifiers and reconcile effects before retrying. An error response does not prove that the external action never completed.

Does passing a 30-day pilot justify unrestricted rollout?

No. Expand only within the tested conditions and approved scope. Changes to workload, integrations, permissions or models need review; continued monitoring and a workable fallback remain necessary.

Download the ratings and full methodology

Download all 50 process ratings (CSV) · Download the complete dataset and methodology package (ZIP).

The release includes each process’s ratings, score, class, suggested approach, human control, validation criterion and rationale. The package preserves the scoring formula, rating definitions and ordering rules. Reproducing the arithmetic verifies consistency with the declared method; it does not validate the judgments or predict business outcomes.

Method, disclosure and corrections

DN developed this editorial framework with AI assistance. The ranking and fictional worksheet are decision aids, not observed performance results, a product endorsement or a certification. The cited sources support particular architecture and control principles; they do not validate DN’s process order.

No affiliate offer or paid ranking appears here. Send corrections through the DN contact page, identifying the process, assumption or source. Include enough non-sensitive evidence to assess the issue. The 2027 address is an editorial series label; the verification date is 22 October 2026 (SAST).

Primary sources

1. Anthropic: Building effective agents, 19 December 2024 — workflow and agent architecture principles; vendor engineering guidance.
2. NIST: Artificial Intelligence Risk Management Framework, AI RMF 1.0, 26 January 2023 — voluntary risk-management framework.
3. OWASP: LLM06:2025 Excessive Agency — functions, permissions, approval and downstream authorization.