What an AI readiness assessment inspects, how we score it, and what you get

An AI readiness assessment is a structured look at how work moves through your company, ending in a ranked register of what to change and in what order. The output is written to be handed to whoever does the building, including a team that is not us.

The scoring model is published below, because it is usually the part a buyer has to take on trust before committing. Assessments run remotely, on evidence you already have, and nothing here needs production access.

Book a readiness assessment

What the assessment answers

Four questions, in this order.

Which repeated work is worth changing at all? Plenty of irritating tasks run four times a year and take twenty minutes. Irritation is not cost.

Which are ready now, and which need the process fixed first? A workflow with an inconsistent starting point cannot be automated usefully. Automating it locks the inconsistency in and hides it.

Where does a mistake cost more than the time saved? Filing a document in the wrong folder is a nuisance. Sending a quote with the wrong price is not.

What is the correct first thing to build? One workflow, with a named owner, a defined failure mode, and a way to tell whether it worked.

How the assessment runs

The shape is the same every time, so you can plan around it.

StageWhat happensTiming
StageScoping callWhat happens20 minutes. You describe the one or two processes you want looked at. If they are a poor fit, that is said here, before anything is bookedTimingBefore you commit
StageEvidenceWhat happensYou send the material listed below. Most of it already exists as a recording, an export or a folderTimingYour pace, usually a few days
StageWorking sessionWhat happens60 to 90 minutes on a screen share. You do the task while we watch and ask questions. Monthly tasks are walked through using the last real runTimingOne session
StageWritten outputWhat happensThe scored workflow list, opportunity register, risk register, first-pilot recommendation and 90-day sequenceTimingWithin five working days of the session
StageFollow-upWhat happens30 minutes to go through it. Optional, and the document stands on its own without itTimingYour pace

What it costs. Priced per assessment as a fixed scope, quoted in writing once the number of processes is agreed, and invoiced by Apefo Ltd. There is no price list on this site. Two assessments that sound identical on a call can differ once you see how the source data behaves. Buying an assessment does not commit you to a build, and the document is yours whether or not you use us for the implementation.

What you own afterwards. The written output, including the mapping of every step and the reasons a workflow was scored the way it was. It is written to be usable by another supplier, or by nobody.

What gets inspected

We ask for evidence, not descriptions. What people say a process is and what the files show can differ, and the gap between the two is where the finding sits.

  • A screen recording of one person doing the task end to endUnedited, including the pauses, the tab switching and the moment they go looking for something. Ten to twenty minutes beats an hour of interviews.
  • Three months of a recurring report, with the source files behind each oneCompared month to month, shifting column orders and hand-corrected cells surface immediately.
  • The CRM field list, marked with which fields are mandatory and which get filled in practice, plus an export of a hundred records. A field that is optional and empty on 80% of records cannot carry routing logic.
  • One week of inbound enquiries and how each was routedWho picked it up, how long it sat, what turned it into a task or a quote, and which ones fell through.
  • The folder where document templates liveAll of it, including final_v3 and final_v3_USE_THIS. Version sprawl predicts where an assistant will later cite the wrong thing.
  • The tool list with who administers each one, by nameA tool with no named administrator is where an integration breaks without anyone being responsible for noticing.
  • Two or three examples of the same task done badlyThe rejected quote, the report that was redone, the enquiry answered twice. Failures describe the boundaries better than successes.

Anything on that list that does not exist is itself recorded. Production access is not needed. Exports, recordings and read-only views cover it, and redacted samples are fine where the shape of the data matters more than the content.

How workflows are scored

Each workflow scores on seven factors, 1 to 5. Four build an opportunity score out of 20. Two build a caution score out of 10. The seventh is a gate, and it carries no score.

FactorWhat is scored15Role
FactorFrequencyWhat is scoredHow often it runs1Quarterly or less5Daily or moreRoleOpportunity
FactorTime per runWhat is scoredHands-on minutes, from the recording1Under 10 minutes5Over 2 hoursRoleOpportunity
FactorPeople involvedWhat is scoredHow many touch it before it is done1One, start to finish5Four or more, with handoffsRoleOpportunity
FactorInput consistencyWhat is scoredHow alike each run's starting material is1Free-text email, no fixed fields5Same form, same fields, every timeRoleOpportunity
FactorError reversibilityWhat is scoredHow hard a mistake is to undo1Undone internally in minutes5Reaches a client or external recordRoleCaution
FactorOutput sensitivityWhat is scoredWhat the output contains or commits to1Internal, non-binding5Personal data, pricing, a commitmentRoleCaution
FactorNamed ownerWhat is scoredWho owns it after handover1Nobody5Named, and they knowRoleGate

The combinations matter more than the numbers:

  • Opportunity 15+, caution 4 or below, owner namedBuild first. Frequent, slow, consistent input, small blast radius.
  • Opportunity 15+, input consistency 1 or 2Fix the intake first, usually by replacing a free-text request with a form that has required fields. It is a small piece of work, and it removes most of the branching the later build would otherwise have to handle.
  • Opportunity 15+, caution 8+Split it. Automate the preparation, leave the decision with a person: draft the quote, do not send it. Anything crossing an external boundary keeps a human on the last step.
  • Opportunity 10 to 14Stays on the register and waits. A first build often removes the dependency suppressing the score.
  • Opportunity below 10Left alone, written down as left alone, with the reason. Low-frequency automations generate maintenance forever and repay nothing.
  • No named owner, whatever the scoreNot built. Unowned automations fail silently when an upstream system changes, and nobody notices until a client does.

Time per run comes from the recording, never an estimate. Estimates go wrong in both directions. Dull tasks get overestimated, and the searching and waiting get left out entirely.

Where the time goes

Recordings are marked up against six bottleneck categories. Expect a single recording to surface two or three of them. Which one dominates decides whether the fix is software, a document convention, or a conversation with a manager.

BottleneckHow it is recognisedWhat it means
BottleneckWaiting on an approvalHow it is recognisedElapsed time far exceeds hands-on time; the recording ends mid-taskWhat it meansA threshold or delegation rule. Software will not shift it
BottleneckRe-entering existing dataHow it is recognisedCopy-paste between two windows; the same value typed twiceWhat it meansAn integration or field mapping
BottleneckSearching for the current versionHow it is recognisedRepeated folder navigation, files opened and closed, a colleague askedWhat it meansA naming convention, before anything is built
BottleneckReconciling two sources that disagreeHow it is recognisedA third spreadsheet exists whose only job is comparisonWhat it meansOne source must be made authoritative; the assessment says which
BottleneckRework from incomplete intakeHow it is recognisedThe task starts, stops, restarts after a clarifying emailWhat it meansRequired fields at the point of capture
BottleneckManual formatting and assemblyHow it is recognisedThe last third of the recording is layout workWhat it meansTemplating or generation, usually the easiest real win

Waiting and searching are not solved by AI, and buying a tool for them wastes money. Re-entry, assembly and inconsistent intake are.

A sample opportunity register

Illustrative, to show the format. The workflows and the scores below are invented.

WorkflowFrequencyWhat changesTypeScoreSequence
WorkflowMonthly client performance reportFrequency12/year, ~90 minWhat changesData pulled and assembled automatically; commentary written by a personTypeAutomation + draftScore16 / 3Sequence1
WorkflowWebsite enquiry to CRM recordFrequency~40/weekWhat changesForm fields mapped to CRM fields; incomplete submissions held in a queueTypeIntegrationScore15 / 2Sequence2
WorkflowQuote preparation from a specFrequency~8/weekWhat changesDraft assembled from the template library; pricing and sending stay manualTypeAssisted draftScore17 / 9Sequence3
WorkflowInternal policy questionsFrequencyDaily, ad hocWhat changesAssistant answers from a signed-off document set, with source linksTypeInternal assistantScore14 / 4Sequence4
WorkflowSupplier invoice codingFrequencyMonthly, ~40 minWhat changesNothing yet. Two systems disagree on supplier namesTypeProcess fix firstScore11 / 6SequenceBlocked
WorkflowAnnual insurance renewal packFrequency1/yearWhat changesNothingTypeNoneScore6 / 7SequenceLeave

The Sequence column is the deliverable. An opportunity list with no order hands the hardest decision, which one to start, straight back to the person who commissioned it.

A sample risk register

Also illustrative. A real register carries only the risks the evidence surfaced.

RiskWhere it shows upWhat we do about it
RiskAn assistant cites a superseded documentWhere it shows upFolders of near-identical versions; policies updated in place, undatedWhat we do about itAssistants read a defined source set, not a whole drive; superseded files move out of scope first
RiskAn automation fires twiceWhere it shows upA retry after a timeout, or two triggers on one event, producing duplicate records or a repeated emailWhat we do about itEach write step carries an idempotency key from the source record; duplicates are rejected and logged
RiskA field mapping silently drops dataWhere it shows upLong free text truncated into a short CRM field; a date parsed wrongly on importWhat we do about itMapping written field by field before the build, then real records compared field by field after
RiskSensitive data leaves an approved systemWhere it shows upClient files pasted into a general chat tool; personal data in a prompt to an external modelWhat we do about itWhich categories go where is decided during the assessment and written into the pilot brief
RiskNobody notices a broken automationWhere it shows upAn upstream API changes; work reverts to manual and nobody says soWhat we do about itFailures notify a named person; every workflow keeps a documented manual fallback
RiskThe person who understood it leavesWhere it shows upOne operator holds the process in their headWhat we do about itHandover notes are part of the build

Choosing the first pilot

The first pilot is chosen to be checkable, not impressive. The conditions below rule candidates out in roughly this order.

It runs often enough to produce a verdict inside a month. A workflow that fires twice in ninety days cannot be judged in ninety days, whatever it scored.

It has a defined wrong answer. For a report, wrong means a figure that disagrees with the source. For enquiry routing, wrong means the record reached the wrong person, or was created without a company name.

It has a manual fallback that still functions, so that if the automation stops the work is still doable the old way.

Where two candidates are close, the one whose systems have a named administrator wins. Waiting on access is the easiest kind of slippage to see coming, because it depends on people who are not in the room when the pilot is chosen.

The first 90 days

This assumes one pilot. Two at once works when they touch different systems and different people, and is a mistake when they do not.

  1. Day 30The pilot runs in production, with the manual fallback still in place for anything it refuses. The mapping document exists and matches what was built. The question at this point is whether the output is correct on every run, checked against the old method. Time saved is a day 90 question. If the output is wrong, day 30 is where the pilot stops.
  2. Day 60Exception paths have been exercised on real cases, including the incomplete form, the duplicate record and the enquiry that fits no category. Failure notifications have fired at least once and reached the intended person. The owner has run the workflow unaided. Item two is scoped in detail, informed by what item one taught.
  3. Day 90The fallback is documented but no longer routine. Handover notes cover what the workflow does, what it will not do, who to contact, and how to switch it off. The register is re-scored, because a first build moves the numbers behind it.

What this assessment will not tell you

  • Hours saved

    It will not tell you how many hours you will save. Anyone quoting a percentage before watching your recordings is quoting someone else's average, and at the scale of one team those averages mean very little. You get measured time per run and a described change, so you can do that arithmetic yourself and check it after day 30.

  • Whether the team adopts it

    It will not tell you whether your team will adopt it. Adoption turns on whether the people doing the work were consulted, whether the change makes their day easier or only makes the company cheaper, and whether someone senior visibly uses it. Those are management questions, and the evidence can only point at them.

  • Legal and compliance

    It will not give you a legal, GDPR, security or compliance opinion. The assessment records where sensitive data appears and where it crosses a system boundary. Whether a given arrangement is lawful is work for a qualified adviser.

  • Software cost

    It will not price your software stack. Licence costs depend on seat counts, contract timing and negotiation, none of which are visible in workflow evidence. Nor does it cover servers, backups, monitoring or where a model is hosted; that is infrastructure work, handled by our sister brand YNVAR.

  • That everything should be automated

    And it will not conclude that everything should be automated. A fair share of any register ends up marked "leave alone", and that column earns its place by keeping effort away from work that will not repay it.

Main entry offer

AI Readiness Assessment

Map the workflow before choosing the tool.

A practical diagnostic for teams that want to reduce manual work without starting from a tool list or a vague AI idea.

Diagnostic path Workflow first

The audit identifies repeated work, tool handoffs, data sensitivity, and the review points that should remain human-owned.

01 Request

Share the workflow, team, tools, and repeated manual work.

02 Map

Trace handoffs, owners, data sources, and review points.

03 Prioritise

Separate process cleanup, rules, AI support, and integrations.

04 Pilot

Choose one realistic workflow to test before expanding.

01 Workflow summary
02 Bottleneck map
03 Repetitive task inventory
04 Tool and data flow notes
05 Automation and AI opportunity list
06 Risk and human review notes
07 Recommended first pilot
08 30/60/90-day roadmap

Diagnostic tool

What the task costs you now

Your own numbers, multiplied. It tells you what a repeated task currently costs in hours and internal cost, which is the figure worth knowing before deciding whether an assessment is worth booking.

It does not estimate what you would save. That number depends on how consistent your inputs actually are, and nobody can give it to you honestly before looking at the work.

Hours per month, currently 0
Internal cost per month 0

Based on 4.33 weeks a month. Hands-on minutes only. If the figure comes from an estimate rather than a recording, expect it to be wrong in both directions.

Book a readiness assessment

FAQ

What if we cannot produce everything on that evidence list?

Then we work with what exists, and the gaps get recorded rather than filled in with assumptions. Two items carry most of the weight. One unedited screen recording, and three months of a recurring report with its source files. With only those, the assessment still runs, and it covers fewer workflows.

Can you do this without screen recordings?

A live walkthrough on a call substitutes, and it is weaker. People narrate the process they believe they follow and skip the searching and waiting, because those parts feel like they do not count. A recording keeps the clock running.

Who needs to be in the room?

The person who does the work, for the walkthrough. The person who owns the outcome, for the scoring decisions. They need not be the same person, and where their two accounts differ is worth pursuing, because a process described two ways is being run two ways.

What if two people describe the same workflow differently?

It gets recorded as a finding rather than settled on the call. One process run two ways has no single definition, which caps input consistency at 2 and pushes it into "fix the process first". Fixable, and worth catching before anything is built on top of it.

Do you need access to our systems?

Not for the assessment. Exports, screenshots, recordings and read-only views are enough. Access matters at build time, and the assessment records which accounts and permissions a pilot needs, so the request does not arrive as a surprise.

Which processes get ruled out, and why?

Workflows get ruled out for reasons a build cannot fix. Nobody owns the process, which stops the work outright. Or it runs two different ways, which caps input consistency and routes it to fix-the-process-first rather than to a build. Which processes you should not automate works through the screen in full, including the eight factors that sit behind the scores above.

What if the conclusion is that we should not build anything yet?

Then that is written up with reasons. The condition that produces it is intake inconsistent enough that any automation would encode the inconsistency. The remedy is usually small, and the register stays valid until it is done.

How much of your own experience sits behind this method?

All of it. The seven factors came out of our own automations. They are the ones that predicted which of them survived real use and which were quietly abandoned inside a year, on internal workflows we run ourselves. The four opportunity factors weigh equally, out of 20. The two caution factors weigh equally, out of 10. The gate is not scored at all. A high caution score does not kill a build. It forces the split, with a person on the committing step. Only a missing owner stops the work.

Next step

Start with the process, not the tool

Describe one task your team repeats and you get a straight view on whether it is worth changing.