How Consultants Assess AI Readiness in Businesses

Consultants assess AI readiness in a business by talking to four kinds of people, asking each of them about the most recent real instance of the work rather than how it usually goes, checking what they're told against the records, and rating each task against a written rubric. The areas they cover (the work, data, systems, people, rules, and direction) are the easy part to describe. What separates a real assessment from a questionnaire is who gets asked, how the questions are put, and what the consultant is listening for in the answers.

That second part is mostly missing from what's published. We read the 14 readable pages ranking for this question and close variants in September 2026. Two published a list of questions, and neither separated them by who gets asked: one was a self-scoring quiz, and the fullest, ConsultKit's ten questions, is framed as a discovery call for qualifying prospects. Three published rubrics with a description for each level, and all three were about data pipelines and model operations, not about the work an office actually does. None gave separate questions for the owner and the person who does the job.

This page fills that gap. It sets out who a consultant should talk to, the questions for each of them, the signals in the work that matter most, and a rubric for rating a task. Then it runs all of it through one worked example, a 60-person machine shop's quoting process. It's the "questions and rubrics" companion to our guide to AI readiness assessment services, which covers what an assessment examines and how the engagement runs. Avolis runs these assessments, so we have a stake in how you judge them, and the last main section is a set of checks you can apply to any consultant, including us.

Key Takeaways

• A consultant should talk to four roles: the owner, whoever runs operations, the person who does the work, and the person who receives what the work produces.
• Of 14 ranking pages we reviewed in September 2026, none gave separate questions for each role, and the only anchored rubrics covered data pipelines, not office work.
• Fixed questions beat conversation. In hiring research, structured interviews predicted job performance with a validity of .42, against .19 for unstructured ones (Sackett and others, 2022).
• Ask about the last time, not the usual time. The method behind that rule, the critical incident technique, dates to 1954 and exists to replace "opinions, hunches, and estimates."
• An LLM alone could speed up about 15% of US work tasks, against 47% to 56% with software built around it, so most of what a consultant rates is the connection, not the model (Eloundou and others).

Table of Contents

• What does a consultant actually do in an AI readiness assessment?
• Who should a consultant talk to?
• Why the questions are fixed and ask about the last time
• The questions, by role
• What do consultants look for in the work itself?
• How does each task get rated?
• A worked example: quoting at a 60-person machine shop
• How to tell whether a consultant assessed or just surveyed
• Where Avolis fits
• Frequently Asked Questions
• Continue Learning

What Does a Consultant Actually Do in an AI Readiness Assessment?

A consultant gathers three kinds of evidence about each workflow in scope: what people say about it, what the records show, and what the work measures when someone counts and times it. Each one checks the others. The owner's account of how quoting works gets compared with the estimator's account of the last quote, and both get compared with how many quotes the system says went out last month.

The methodology page ranks these by how much they prove, in its section on what counts as evidence. Here the point is simpler: a consultant needs all three, because each one catches what the others miss.

The steps an engagement runs through (scope, measure, score, decide) are laid out in the pillar guide's section on how assessments run, and the list of what to have ready is in what to prepare. How the ratings combine into a decision is covered in our AI readiness assessment methodology. This page stays on the fieldwork itself, which is where most of the judgment happens and where you can most easily tell a careful assessor from a hurried one.

Who Should a Consultant Talk To?

A consultant should talk to four roles for each workflow: the owner, whoever runs operations, the person who does the work, and the person who receives what the work produces. At 10 to 200 people that's usually a handful of conversations, because one person often wears two of the hats, and it's rarely a round of interviews with every department.

• The owner knows what the business is trying to do more of, which risks they won't take, and what's been tried before. They usually don't know how long a task takes, and they shouldn't be expected to.
• Whoever runs operations knows how the pieces connect, which system wins when two disagree, and where the counts live. In a small firm that's the office manager, the operations manager, or sometimes the owner's spouse who does the books.
• The person who does the work is the only one who can walk through the last instance step by step. This is the estimator, the dispatcher, the coordinator who books the vendor.
• The person who receives the output is the one most assessments forget. It's the production planner who turns a won quote into a job, the bookkeeper who gets the invoice, or the customer.

One security consultancy's guide makes the point about the third role plainly: the interview with someone who does the work the AI is meant to change "is routinely skipped and routinely the most useful" (Purple Shield Security, AI Readiness Assessment Consultants). We'd add the fourth role for a reason that comes from research, not preference. One of the most widely cited studies of which tasks AI can speed up defines "equivalent quality" as output where "a third party, typically the recipient of the output, would not notice or care about LLM assistance" (Eloundou, Manning, Mishkin, and Rock, GPTs are GPTs, working paper, 2023). If the quality test belongs to the recipient, the recipient has to be asked.

Change readiness across a larger organization, meaning leadership alignment, manager support, and how teams take to new ways of working, is a bigger subject with its own methods, and it's covered in how consultants assess AI readiness in organizations. At this size, the people question is narrower: will the person who does the work use what gets built, and who will own it when something breaks?

Why the Questions Are Fixed and Ask About the Last Time

Good assessment questions are written down in advance, asked the same way every time, and aimed at a specific recent instance of the work. Both habits come from decades of research on interviewing, and both exist for the same reason, which is that a free-flowing conversation about how things usually go produces confident answers that don't hold up.

The clearest evidence for fixed questions comes from hiring, which is the most studied interview of all. In a 2022 reanalysis of the major meta-analyses of selection methods, structured interviews, where every candidate gets the same questions and answers are rated against the same standard, came out as the top-ranked method, with a mean validity of .42. Unstructured interviews came in at .19 (Sackett, Zhang, Berry, and Lievens, "Revisiting Meta-Analytic Estimates of Validity in Personnel Selection", Journal of Applied Psychology, 2022). Validity here is a correlation between what the interview predicted and how the person went on to perform, so the structured version was more than twice as informative.

A readiness interview isn't a hiring interview, and we're not claiming the numbers transfer. The mechanism does, though. When two assessors ask different questions of different people, their findings differ because of the questions, and the owner can't tell which difference is real.

The second habit, asking about the last time, is older still. In 1954 the psychologist John Flanagan described the critical incident technique, a method of collecting accounts of specific things people actually did. He recommended tying each question to "the most recent observation," so that people don't report only "the more dramatic or vivid incidents" they happen to remember. His summary of the method's purpose works as a summary of good readiness fieldwork, too: rather than collecting "opinions, hunches, and estimates," it records specific behavior from the people best placed to observe it (Flanagan, "The Critical Incident Technique", Psychological Bulletin, 1954).

That last row is worth taking seriously, and even sales-minded guides agree. ConsultKit's question list, written for consultants qualifying prospects, warns that "You have good data practices, right?" is "not a discovery question" (ConsultKit, AI Readiness Checklist). The gap between what people report and what the records show is real even in the best data. A study linking Danish surveys to payroll records found that workers reported productivity benefits from AI chatbots, while their recorded hours and earnings showed precise null effects two years after ChatGPT's launch (Humlum and Vestergaard, Still Waters, Rapid Currents, NBER Working Paper 33777, revised March 2026). Neither account was dishonest. They were measuring different things, which is exactly why an assessor needs both.

The Questions, by Role

The questions below are the core set we'd expect a careful consultant to cover, grouped by who gets asked. Each one is tied to a recent instance or to something that can be checked. The notes say what each answer is for. You can use them to prepare, to judge a consultant's interviews, or to run a rough version yourself.

The Owner

1. "A year from now, what do you want the business doing more of?" This sets direction. AI that serves a goal gets ranked above AI that is the goal.
2. "Whose vacation worries you most, and why?" It finds the work that lives in one person's head faster than any question about "key-person risk."
3. "What have you already tried with AI or new software, and what happened to it?" The answer shows which failure the business is most likely to repeat.
4. "Which customers, contracts, or insurers limit where your data can go?" Most owners know the answer for their biggest customer and haven't checked the rest.
5. "What would make you call this a waste of money in six months?" It turns a vague hope into a measure, and it's the measure the result should be judged by.

Whoever Runs Operations

1. "Where would I find last month's count of [quotes, work orders, intakes]?" If the answer is "I'd have to estimate," that's the first finding.
2. "When two systems disagree, which one wins?" Every business has a source of truth, and it's often a spreadsheet rather than the system everyone pays for.
3. "What's the most common reason one of these gets sent back or redone?" Rework is where a lot of the time goes, and it rarely shows up in anyone's estimate.
4. "Which steps exist because of one customer, one past mistake, or one person's preference?" Those are the steps to question before anyone automates them.
5. "Who could change how this workflow runs, without asking anyone?" If nobody can, the workflow has no owner, and nothing built on it will stick.

The Person Who Does the Work

1. "Take me through the last one you did, start to finish." This is the most important question on the page. Everything else follows from it.
2. "What did you look up, and where did you find it?" The answer maps where the knowledge lives, which is usually more places than anyone expected.
3. "What did you type that already existed somewhere else?" Retyping is one of the easiest things for a build to remove, and people rarely mention it unprompted.
4. "How long did that one take, and was it typical?" Asked right after the walk-through, the time estimate is anchored to a real instance instead of a feeling.
5. "What do you do when one doesn't fit the usual pattern, and how often does that happen?" Exceptions decide how much of the work can be handed off and how much still needs a person.
6. "Which part would you never want to hand to anyone else?" It usually marks the judgment the job depends on, and it tells the consultant what the build has to leave alone.

The Person Who Receives the Output

1. "What do you check when one of these reaches you?" This is the quality standard, and it's often unwritten.
2. "What's the most common thing wrong when it arrives?" It shows where the errors come from, which may not be where the owner thinks.
3. "If it were prepared a different way but came out the same, would you notice or care?" That's the equivalent-quality test from the research above, asked of the person it belongs to.

None of these questions mentions AI until the owner's third one, and that's deliberate. People describe their work more accurately when they aren't guessing what the consultant wants to automate. For the owner-side version of this, written as a self-check rather than an interview, see how to assess your organization's AI readiness yourself.

What Do Consultants Look For in the Work Itself?

While the person walks through the last instance, a consultant is listening for a short list of signals, and each one points to a different kind of finding. None of them is good or bad on its own. They tell the consultant where to look next, and together they decide whether a task is worth building for.

The signal that matters most: "ask [one person]" is where most promising workflows stall, and recent usage research points the same way. Anthropic's analysis of how businesses deploy its models concluded that AI for complex tasks "might be constrained more by access to information than on underlying model capabilities," especially where "tacit, diffuse knowledge" runs the operation (Anthropic, Economic Index: Uneven AI Adoption, September 2025). In a 60-person business, that tacit knowledge usually belongs to one estimator or one dispatcher. An assessment that doesn't find out what that person knows, and where it could come from instead, hasn't assessed the workflow.

Two signals on the list point away from AI altogether. Long waits usually mean the fix is a process change, like a daily cutoff or a reminder to the customer. A workaround for a tool you already pay for means someone should turn on the feature. A consultant who reports those as findings, rather than dressing them up as AI projects, is doing the job properly. Our guide to why AI projects fail traces how often the wrong target, not the wrong model, is what sinks a project.

How Does Each Task Get Rated?

Each task gets rated against a written rubric, meaning a short description of what a weak, middling, and strong answer looks like for each question, so that two assessors hearing the same answers land in the same place. The rubric below covers the work itself, which is the area the anchored rubrics we found online left out.

The first question a consultant asks of any task is whether AI could make it meaningfully faster without making it worse. The research that defined that question most carefully set the bar at half: a task counted as exposed if a language model could cut the time it takes "by at least half (50%)" at equivalent quality (Eloundou and others). On that test, the authors estimated that a language model on its own could speed up about 15% of all US worker tasks, and that software built around the model raised the share to between 47% and 56% (Eloundou and others, working paper, 2023; the peer-reviewed version appeared in Science in 2024).

For a business owner, the gap between those two numbers is the whole point. Most of the gain doesn't come from the model answering questions. It comes from connecting it to the drawings, the job records, and the systems the work already runs on, which is why an assessment spends so much time on where information lives and so little on which model to use. The authors are also candid that they expected real-world savings "would likely be slightly or significantly lower" than their estimates (working paper), so a consultant should treat the half-time test as a filter, not a forecast.

Once a task passes that first filter, it gets rated on six questions about the work. The first three are the gate our methodology describes, which asks whether the task happens often enough to matter, whether someone can say roughly how long it takes, and whether one person owns it. So a weak rating on any of the three sends the task to the wait list with the reason written down, and so does a middling one on ownership. The other three questions shape how much effort the build will take.

A rubric like this is only as good as its descriptions, which is why each cell describes something you could verify. "Strong" on the time question doesn't mean a confident estimate. It means someone timed it. The same logic runs through our readiness checklist, where each question says what counts as a yes.

A Worked Example: Quoting at a 60-Person Machine Shop

Here's how the questions, signals, and rubric come together on one workflow. The business is a composite we've put together for illustration, not a client: a 60-person precision machine shop that makes parts to customers' drawings, and whose owner believes quoting is slowing the business down. All the figures below are illustrative.

The owner says quoting "takes an hour or so" per request and that the shop loses work because quotes go out late. Asked whose vacation worries them most, the owner names the senior estimator without hesitating. Asked what would make this a waste of money, the answer is concrete: if quotes still take more than two days to go out.

The operations manager pulls last month's count from the shop's ERP system, the software that runs orders, jobs, and inventory. There were 140 requests for quote, 41 of which became orders. Asked who could change how quoting runs without asking anyone, the manager names the senior estimator. Asked which system wins when two disagree, the manager admits it's the estimator's spreadsheet of past jobs, which holds the actual machine hours that the ERP system often doesn't.

The estimator walks through the last three quotes, starting with the most recent, which arrived as an email with a PDF drawing attached. The estimator read the drawing, typed the part number, material, and quantities into the ERP system, emailed a supplier for a material price, searched the spreadsheet for similar parts to estimate machine time, and built the quote. That one took about 50 minutes of work spread across two days, most of the gap spent waiting on the supplier, and the other two took 35 and 70 minutes. The estimator puts roughly a third of requests in the "weird" pile, meaning tight tolerances or unusual materials, and says the similar-parts judgment is the one piece they'd never hand off.

The production planner, who receives every won quote, says the most common problem is a quote that assumed a machine the shop has since stopped using. Each won quote also gets retyped a second time into a job traveler, the document that follows a part through the shop.

The signals are easy to spot once you know the list: the same drawing data typed twice, a spreadsheet outranking the ERP system, one person holding the judgment that matters most, and a two-day wait that has nothing to do with AI. A plausible result splits the workflow in two. Reading the drawing and filling in the ERP system and the job traveler is a go, because it's frequent, measured, owned, and mostly retyping. Suggesting machine time from past jobs is a wait until the spreadsheet's actual hours are consistently recorded against each job. The supplier wait is a process fix, such as a standing price list for the materials the shop buys most, rather than an AI project. The pillar guide's decision table explains what each of those outcomes commits you to.

Citation-ready summary: A consultant assessing AI readiness in a business should interview four roles for each workflow (the owner, whoever runs operations, the person who does the work, and the person who receives the output), ask each about the most recent real instance rather than the usual case, check the answers against system counts, and rate each task against a written rubric for frequency, time, ownership, variation, where the knowledge lives, and who checks the result.

How to Tell Whether a Consultant Assessed or Just Surveyed

You can tell whether a consultant assessed your business or just surveyed it by looking at what they asked for and who they talked to, before you ever read the report. These checks apply to any consultant, including us, and each can be answered by the end of the first week.

If a consultant fails more than one or two of these, the report may still be tidy, but it describes your business as you described it to them, which you already knew. Pricing, provider types, and what the finished assessment should leave behind are covered in our comparison of AI readiness assessment consulting firms, and the wider choice of partner starts with which AI consulting company to choose.

Where Avolis Fits

The questions and rubric on this page are how our two-week diagnostic works. We work through each workflow in scope with the people who run it, starting with the last real instance, then pull the counts from your systems and time the work. Each task is rated against a written rubric, and you get a ranked result with a go, wait, or no on each workflow. Every engagement with us starts with that diagnostic, and the ranked result is yours whatever you decide to do next.

We wrote these criteria from how we run our own assessments, so weigh them with that in mind. If you'd rather work through the questions yourself first, how businesses assess readiness for AI adoption covers the internal version.

Frequently Asked Questions

How do consultants assess AI readiness in a business?

They interview the owner, whoever runs operations, the person who does the work, and the person who receives it, asking each about the most recent real instance. They check the answers against system records, count and time the work, and rate each task against a written rubric, ending in a go, wait, or no decision on each workflow.

What questions do consultants ask in an AI readiness assessment?

The most useful ones are tied to real, recent work: "Take me through the last one you did," "What did you type that already existed somewhere else?", "Where would I find last month's count?", and "Whose vacation worries you most?" Questions about opinions or ratings, such as "Is your data in good shape?", produce the least reliable answers.

Who should a consultant interview for an AI readiness assessment?

A consultant should interview four roles for each workflow: the owner, whoever runs operations, the person who does the work, and the person who receives the output. At 10 to 200 people that's usually a handful of conversations. The person doing the work is the interview most often skipped, and the one most likely to show where time actually goes.

How is a consultant's assessment different from a self-assessment?

A self-assessment records how you describe your business. A consultant's assessment checks that description against records and measurements: it counts last month's volume, times recent real instances, and hears from the people who do and receive the work. The difference shows up most in how long tasks take and where the knowledge lives.

How can you tell if an AI readiness assessment was done properly?

Check whether the consultant talked to the person who does the work, asked for counts and real examples rather than ratings, timed anything, and can show their questions and rubric in writing. A thorough assessment also reports at least one finding that isn't an AI project, such as a process fix or a feature you already pay for.

Continue Learning

The areas an AI readiness assessment covers are the same almost everywhere. What differs is whether anyone asked the estimator about the last quote, and whether the answer was checked against a count.

Before you hire anyone, ask to see the questions they'll ask your team.

How assessments work:

• AI readiness assessment services
• AI readiness assessment methodology
• How consultants assess AI readiness in organizations
• AI readiness assessment checklist and PDF

Assess it yourself:

• How businesses assess readiness for AI adoption
• How can I assess my organization's AI readiness?
• AI readiness assessment for SMBs

Choosing who does it:

• AI readiness assessment consulting firms
• Which AI consulting company should I choose?
• Why do AI projects fail?

---

Sources

All sources retrieved 2026-09-24.

• Tyna Eloundou, Sam Manning, Pamela Mishkin, and Daniel Rock, GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models, arXiv:2303.10130, working paper, August 2023 revision; published as "GPTs are GPTs: Labor market impact potential of LLMs," Science 384(6702), 1306–1308, 2024, retrieved 2026-09-24 — https://arxiv.org/abs/2303.10130
• Paul R. Sackett, Charlene Zhang, Christopher M. Berry, and Filip Lievens, "Revisiting Meta-Analytic Estimates of Validity in Personnel Selection: Addressing Systematic Overcorrection for Restriction of Range," Journal of Applied Psychology 107(11), 2040–2068, 2022, retrieved 2026-09-24 — https://pubmed.ncbi.nlm.nih.gov/34968080/
• John C. Flanagan, "The Critical Incident Technique," Psychological Bulletin 51(4), 327–358, 1954, retrieved 2026-09-24 — https://www.apa.org/pubs/databases/psycinfo/cit-article.pdf
• Anders Humlum and Emilie Vestergaard, Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI, NBER Working Paper 33777, May 2025, revised March 2026, retrieved 2026-09-24 — https://www.nber.org/papers/w33777
• Anthropic, Anthropic Economic Index: Uneven AI Adoption, September 15, 2025, retrieved 2026-09-24 — https://www.anthropic.com/research/anthropic-economic-index-september-2025-report
• Purple Shield Security, AI Readiness Assessment Consultants, retrieved 2026-09-24 — https://www.purpleshieldsecurity.com/post/ai-readiness-assessment-consultants
• ConsultKit, AI Readiness Checklist: 10 Questions to Ask Every Client Before Starting, March 2026, retrieved 2026-09-24 — https://www.consultkit.ai/blog/ai-readiness-checklist-10-questions-to-ask-every-client-before-starting-1773254396052

On the page review. "The 14 readable pages" are the pages that described how a consultant or firm runs a business AI readiness assessment among the results for "how consultants assess ai readiness in businesses," "how do consultants assess ai readiness," "ai readiness assessment interview questions," and "ai readiness scoring rubric" on 2026-09-24: ConsultKit, Sketch Development, Purple Shield Security, Shispare, RSM US, Centric Consulting, Third Stage Consulting, Panorama Consulting, Thinklytics, OvalEdge, Saigon Technology, Paiteq, Intuz, and Summit Trails. ConsultKit and Sketch Development published question lists (Sketch's is a self-scoring quiz). OvalEdge, Saigon Technology, and Intuz published rubrics with level descriptions, all for data, pipelines, or model operations. SkillPanel's page was blocked to automated retrieval and isn't counted. It's a snapshot of one day's results, not a market survey, and the commercial pages are cited for their methods, not as evidence of anything else.

On the hiring research. The Sackett figures are validities for predicting job performance in employee selection. We cite them for what they show about structured and unstructured interviews in general, not as a measure of readiness assessments, which have no comparable body of research.

On the task-exposure estimates. The 15% and 47–56% figures come from the 2023 working paper, which reports them as shares of all US worker tasks under the authors' rubric. The Science version reports its results by job rather than by task. Both are estimates of what's technically possible, not of what firms have adopted, and the authors expected real-world savings to be lower.

On first-party claims. Statements such as "our two-week diagnostic" describe Avolis's own engagements. They are not independent research and are not offered as benchmarks. The machine shop is an illustrative composite, not a client, and its figures are illustrative.

---

About Avolis Research Group

Avolis Research Group is Avolis's in-house research practice, focused on how operations-heavy small and mid-sized businesses actually adopt AI. It synthesizes primary economic research, government survey data, and results from real implementations into practical, vendor-neutral guidance.

More about Avolis and how we work · Get in touch