Your AI assessment should end with a workflow, not a shopping list

Most AI assessments for accounting firms start with the wrong question: "Which tools should we look at?"

That feels productive for about three minutes. Then the demo ends, the partner meeting starts, and the same ugly work is still sitting there. Source documents arrive in five flavors of incomplete. Staff notes live in one place, client answers in another, and the review comments explain the real process better than the procedure manual. One exception turns a clean automation idea into another file someone has to reopen on Friday afternoon.

That is why Intelligence Solved treats assessment work as operator triage. The job is not to widen the menu of AI possibilities. The job is to narrow the mess until one accounting workflow is safe, bounded, and useful enough to install.

Problem

Tool shopping is usually procrastination with a better interface

The accounting AI conversation is full of examples: document intake, workpapers, tax prep, extraction, reconciliation, continuous close, exception review. Those are legitimate areas to examine. They are also terrible starting points if the firm has not decided which workflow can survive its actual operating conditions.

A tool-first assessment lets everyone feel modern without forcing a hard operational decision. The firm gets a list of options. Maybe a readiness score. Maybe a slide that says the team should "prioritize automation opportunities." Then everyone goes back to the same client email threads, the same late uploads, the same unclear handoffs, and the same reviewer who will not sign off because the output does not match how the firm actually works.

The painful part is not that the firm lacks AI ideas. Most firms have too many. The painful part is that every idea has to pass through the same narrow doorway: real documents, real people, real review standards, real exceptions, and real accountability.

An assessment that skips that doorway is not strategy. It is a nicer way to postpone the decision.

First filter

The first useful question is: which workflow can hold still long enough to install?

A responsible assessment starts by reducing the target.

Not "Where can AI help the firm?" That question is too big. It invites abstract answers and vendor-flavored wish lists.

A better question is: "Which one workflow creates enough pain to matter, has enough repeatability to assess, and has boundaries clear enough that we can define what good looks like?"

For an owner-led bookkeeping, CPA, or CAS firm, that distinction matters. The owner is not buying an idea. The owner is carrying the cleanup when the idea breaks. Reopened work. Review bottlenecks. Client document chaos. Staff uncertainty. Partner hesitation. The quiet embarrassment of having to explain why the new tool made the work look faster online but messier inside the firm.

So the assessment has to find a workflow that can be named without hand waving. A month-end close handoff. A document intake lane. An exception review path. A tax prep source-document trail. Whatever the first installable workflow is, it must be specific enough that the firm can say what enters it, who touches it, what breaks it, who reviews it, and what evidence would prove it worked.

If the workflow cannot be named that clearly, implementation is premature.

Evidence

The assessment is where the operating evidence gets pulled out of hiding

Accounting firms often know more about their broken workflows than they think. The evidence is just scattered across the work.

It is in the portal uploads that still require follow-up emails. It is in staff notes that explain the exception better than the system does. It is in review comments that appear every month because the upstream handoff never changed. It is in the client answers that arrive too late, too vague, or in the wrong format. It is in the owner's head, which is a dangerous place for a critical workflow to live if the goal is to install anything repeatable.

Intelligence Solved assessments turn that scattered evidence into an installation target.

That means naming the workflow boundary first. Where does the work start? Where does it stop? Which documents count? Which handoffs matter? Which exceptions are normal enough to design for, and which ones should still be escalated to a human?

Then the assessment has to expose the control model. Who owns the workflow? What can software draft, sort, extract, summarize, or route? What must a human still approve? What should trigger review? What should never be automated because the judgment risk is too high or the source evidence is too weak?

This is the part generic readiness scorecards tend to miss. A score can say the firm is "ready" or "not ready." It cannot tell the owner which workflow has a clean enough boundary, enough evidence, and enough human control to install first.

Controls

Human controls are not a compromise. They are what make the install possible.

A lot of AI advice quietly treats human review as a drag on automation. In accounting work, that is backwards.

Review standards are not bureaucracy. They are the reason the work can be trusted. If an assessment does not define the reviewer's standard, the install has no finish line. The tool may produce something. The team may like the interface. But the owner still has to answer the question that matters: would we trust this inside our actual workflow?

That requires plain boundaries.

What source trail does the reviewer need? What exception types must be surfaced instead of hidden? What client behavior will break the workflow? Which handoff creates the most rework? Where does the system need to show its work because a confident summary without evidence is worse than no summary at all?

The assessment earns its keep when it makes those questions concrete before the firm buys or builds around the wrong assumption.

This does not make the process slow. It makes it honest. The fastest way to waste time is to automate a workflow nobody has actually defined, then spend the next month cleaning up the edge cases everyone pretended were rare.

Outcome

One installable workflow beats a broad AI roadmap

The outcome of a good assessment should feel narrower than the owner expected.

That is the point.

A broad AI roadmap can sound impressive and still leave the firm unsure what happens Monday morning. A tool shortlist can look useful and still avoid the harder question of whether any selected workflow has enough structure to support installation. A generic AI-readiness score can make the firm feel measured without telling anyone what to fix first.

Intelligence Solved assessment work is designed to reach a different conclusion: one scoped workflow install.

That conclusion should include the boundary of the workflow, the evidence it depends on, the documents and handoffs involved, the exceptions that need a path, the human controls that protect review quality, and the success criteria that tell the firm whether the install is working as intended.

The authority is in the narrowing. Not because tools are unimportant, but because tools only become useful once the workflow has a real shape.

If the firm cannot explain the workflow, the tool will inherit the mess. If the firm can explain the workflow, the first install has something to aim at.

Next move

Start with the workflow that keeps coming back to the owner

The practical next move is not to ask, "What is our AI strategy?"

Start with the workflow that keeps returning to the owner or managing partner after everyone thought it was done.

The file that gets reopened. The client document chase that never quite ends. The review note that shows up every month. The handoff staff avoid because nobody agrees what complete means. The exception path that depends on one person remembering the workaround.

That is where an assessment can become useful.

Do not score the firm in the abstract. Do not turn the exercise into a tour of tools. Put one painful workflow on the table and ask whether it can be bounded, evidenced, controlled, reviewed, and measured well enough to install.

If it can, the firm has a real implementation target.

If it cannot, that answer is useful too. It means the first job is not automation. The first job is making the workflow concrete enough that automation will not just make the mess faster.

Intelligence Solved creates assessments for that decision. One workflow. Clear boundary. Real evidence. Human controls. Review standards. Success criteria. Then, and only then, does implementation have a fair shot at surviving contact with the firm.