AI implementation firms · Opportunity qualification

How AI Implementation Firms Can Win Better Projects Before Writing a Line of Code

How AI implementation firms can qualify opportunities, test the economics and recommend the right intervention before committing delivery resources.

Leania · For AI implementation firms

QUALIFY FIRST

Diagnose the operational problem before choosing the build.

Illustrative diagram. It describes no client and carries no measured figures.

A client asks your team to build an AI agent. The conversation moves quickly to models, integrations, orchestration and technical feasibility. Your engineers can probably build what has been requested, but what problem is the client actually trying to solve?

Perhaps they want to reduce processing costs, clear a backlog or improve service levels. An AI agent might help, but equally the biggest gain could come from fixing incomplete inputs, removing an unnecessary approval or changing how exceptions are handled.

I've seen this across transformation programmes: the solution a business first asks for is not always the intervention that delivers the most value.

For AI implementation firms, that creates an important commercial decision. How do you qualify a client request before committing expensive technical capacity?

A technically feasible project can still be a poor investment

Technical feasibility matters: can the model interpret the documents reliably, integrate with the client's systems and meet their security and control requirements?

Those answers tell you whether something can be built, not whether it should be.

Take an accounts payable team that asks for AI document processing. A closer look might reveal missing purchase orders, inconsistent supplier records and approvals sitting with the wrong person. Better extraction would speed up one step and could still leave the main delay untouched.

So before you design the solution, find out where the effort and the waiting actually occur.

Start with the operational problem

A useful qualification conversation establishes four things.

  • What is happening: the workflow in question, and what its current shape costs the business.
  • How much it matters: volumes, handling time, rework, delay, errors and cost.
  • Why it happens: the root causes, kept apart from the symptoms people notice first.
  • What improvement looks like: an agreed baseline, a target and an accountable owner.

You do not need perfect evidence at this stage, only enough to tell a credible opportunity from an attractive assumption.

Finding where value leaks across a whole operation is a wider exercise, and I've written about it in where I'd look for hidden value before proposing AI. The question here is narrower: does this one request deserve your engineering time?

Six interventions worth considering before choosing AI

Once the problem is clear, compare the realistic options. I use a simple ladder: Stop → Simplify → Standardise → Workflow → Automate → AI.

Six interventions, judged against one outcome

Lower build and running complexity
  • Stop

    Does this need doing at all?

  • Simplify

    Can it take fewer steps, approvals or handoffs?

  • Standardise

    Is the variation justified?

  • Workflow

    Are routing, ownership and escalation clear?

  • Automate

    Are the rules stable and repeatable?

  • AI

    Is the input genuinely variable or unstructured?

Higher build and running complexity

Every option is measured against the same business outcome

A decision framework, not a mandatory sequence. Options can be combined, and none has to be completed before another is considered.

The ladder is a decision framework, not a mandatory sequence. You do not have to finish one rung before you consider the next, and the best answer is often a combination.

  • Stop: remove tasks, checks or reports that add no value, within contractual and regulatory controls.
  • Simplify: cut the steps, approvals and handoffs that slow the work down.
  • Standardise: remove variation that has no good reason, and keep the differences that genuinely matter.
  • Workflow: improve routing, ownership, escalation and visibility.
  • Automate: put stable, repeatable rules into software.
  • AI: interpret variable inputs or unstructured data, where the benefit exceeds the cost of building, running and controlling it.

The discipline is to compare every option against the same business outcome. If the target is a shorter invoice cycle, judge each intervention by how far it moves that number and what it costs to get there.

When the right recommendation is less technology

I've seen this in practice. In an earlier improvement engagement, the recommended intervention was upstream training rather than further automation. The approved business case forecast approximately £150,000 in annual value against around £5,000 in annual running costs.

Those were forecast figures, not independently verified realised savings. The lesson is about diagnosis: finding the correct root cause mattered more than the technology available, and it took some commercial courage to recommend the simpler option.

For an implementation firm that can feel uncomfortable, because a training recommendation is a smaller invoice than a build. Yet advice that fits the problem is how you earn the trust of the people who sign off the next decision.

Put the economics against the actual problem

Here is a synthetic example, which describes no client and uses round numbers to show the arithmetic.

Suppose an accounts payable team processes 50,000 invoices a year, and 30% become exceptions. That is 15,000 exceptions, and if each one needs 10 extra minutes, the team spends 2,500 gross hours a year handling them.

At an illustrative fully loaded cost of £35 an hour, that is £87,500 of gross annual handling capacity cost.

Synthetic example · Not client data, a forecast saving or cash

Gross annual handling capacity cost of invoice exceptions

50,000

Invoices a year

30%

Become exceptions

10 min

Extra minutes per exception

£35/h

Illustrative fully loaded cost per hour

  1. Exceptions a year

    15,000

    50,000 × 30%

  2. Gross handling hours a year

    2,500 h

    15,000 × 10 min ÷ 60

  3. Gross annual handling capacity cost

    £87,500

    2,500 h × £35/h

Not yet known, so no saving is stated

  • What causes the exceptions, from a sample of real cases
  • The proportion that could be prevented
  • The cost to build and run each intervention
  • Whether released time becomes cash or is redeployed
Gross capacity cost measures the size of the problem. A saving needs the four answers above, net of the intervention’s own cost.

That figure is not a forecast saving, and it is not cash; it measures how much capacity the problem consumes today.

To turn it into a business case, you need to know what causes the exceptions in a sample of real cases and what share could be prevented. You also need the cost to build and run each intervention, and to know whether released time would turn into cash or be redeployed to other work.

Until you have those answers, the honest statement is the size of the problem, with its assumptions shown.

Five tests before committing engineering capacity

Before you put a solution architect on the account, run five tests: problem, value, ownership, feasibility and measurement. The scorecard below sets out the question behind each one, applied to the synthetic invoice example.

Synthetic example · An illustrative judgement, revised as evidence arrives

Qualification scorecard: the invoice exceptions request

Qualification scorecard: the invoice exceptions request
ProblemIs there a defined operational issue?30% of invoices become exceptions and need extra handling.Evidenced
ValueAre the economics material?£87,500 gross a year. The preventable share is unknown.Partly evidenced
OwnershipIs there an accountable sponsor?Assume a named accounts payable owner has agreed to sponsor the work.Evidenced
FeasibilityAre data, systems, controls and budget in place?Exception reasons are not recorded consistently. No budget is agreed.Not yet evidenced
MeasurementIs there a baseline and an agreed target?The exception rate is known. No target has been agreed.Partly evidenced

Outcome

  • Proceed

    All five tests hold. Scope delivery against the baseline.

  • InvestigateSynthetic outcome

    A gap could be closed. Run a bounded assessment first.

  • Stop

    No material problem, owner or way to measure it.

Weak evidence leads to a bounded assessment, not automatic rejection.

Each review leads to one of three outcomes: proceed when the evidence is strong, investigate when an answer is missing but could be found, and stop when the problem is minor, has no owner or cannot be measured.

Treat the result as a judgement and revise it as evidence arrives. Weak evidence does not mean automatic rejection; often the right next step is a bounded assessment.

What the client should receive

Qualification should end with a concise, decision-ready brief. It should cover:

  • The problem, its owner and the current baseline.
  • The estimated value, with every assumption stated.
  • The evidence gaps that remain.
  • The intervention options considered, and why one is recommended.
  • The cost, risks and dependencies.
  • The next decision, and the conditions under which the client should reject the proposal.

A brief like this is useful to everyone involved. Your commercial team can price it, your architects can see the constraints and your delivery team knows what success means. The client's finance team can also test the numbers before anyone commits budget.

A better way to grow an AI implementation firm

Better qualification can reduce the number of poorly scoped commitments your firm takes on. It can also change the conversation with clients, because diagnosing the operational problem first often uncovers opportunities beyond the original technical brief.

None of that should be promised without evidence, and the value comes from doing it consistently. When every consultant asks the same questions and applies the same tests, your firm's judgement becomes repeatable across accounts. It stops depending on who happened to run the first meeting.

Start with one live opportunity

Pick one request in your pipeline today and ask a single question.

What problem are we solving, how much is it worth, and what is the simplest credible way to improve it?

If the answers are unclear, a short assessment may be more appropriate than a development proposal. That is still valuable work, and it gives the client a sound basis for the decision that follows.

For AI implementation firms

Want to apply the same discipline to one client account? The free Client Opportunity Expansion Blueprint turns the friction you have observed into ranked opportunity hypotheses, with the unknowns shown, and a client-ready PDF. It needs no sign-in or email.

Build your opportunity blueprint

That's the kind of problem AI-PEX is designed to help implementation firms diagnose, quantify and address consistently. Explore AI-PEX

← All insights