AI implementation firms · Client opportunity expansion

Your Next AI Project Is Probably Already Inside an Existing Client

After an AI document-processing build goes live, the rework, queues, duplication and specialist bottlenecks around it often remain. How to find them, size them honestly and turn them into the next decision your client can take.

Leania · For AI implementation firms

THE NEXT PROJECT

Look upstream and downstream of what you delivered.

Illustrative workflow. The effort around a delivered build is a hypothesis to test with case evidence, not a finding about any client.

You delivered the document-processing solution. It reads the incoming forms, extracts the fields and passes clean data downstream. The client is pleased, the demonstration went well and extraction accuracy is better than anyone expected.

Then you look at the operation around it, and much of the old effort is still there.

Cases still arrive without a signature or a supporting schedule. Someone still repairs the ones the model could not read. Exceptions still wait for the one person who knows the rules, and finance still reconciles the output against another system at month end. The extraction got faster. The workflow around it barely changed.

For an AI implementation firm, that is not a failure. It is often the best-qualified pipeline you will ever have. You already understand the client's data, you have earned some trust and you can see the work from the inside. The question is whether you can turn what you see into a properly evidenced proposal, rather than another speculative pitch for more AI.

What a successful build often leaves behind

Where effort remains after the build

  1. 01 · Upstream

    Information arrives incomplete

    Rework

  2. 02 · Your build

    AI document processing

    Working as designed

  3. 03 · After extraction

    Exceptions wait for an experienced reviewer

    Queues · Specialist bottlenecks

  4. 04 · Downstream

    Data checked again before finance trusts it

    Duplication

Illustrative workflow. The four patterns overlap; one case can be repaired, wait and be reconciled. Count its effort once.

When I look at a workflow after an automation project has gone live, the remaining effort usually falls into four patterns. They are worth naming separately, because each one points to a different kind of fix.

Rework: work that goes backwards. A case is processed, found incomplete, sent back and handled again. The model may be reading the documents perfectly; the documents were simply never complete. The signal is the same case touched more than once for the same outcome.

Queues: work that waits. An exception takes a few minutes to resolve but sits for days before anyone looks at it, because ownership is unclear or approval sits with a senior reviewer. Faster extraction does nothing for a queue created by a rule about who decides.

Duplication: work done twice. The extracted data lands in the case system and is then keyed, or checked, again in a finance or reporting system. Each team has a sensible reason for its own check. Together they pay for the same assurance twice.

Specialist bottlenecks: work that depends on a few people. Routine cases flow through your solution while the awkward ones queue with two or three experienced staff who carry the unwritten rules in their heads. Headline throughput improves; the backlog does not.

These patterns overlap. A rejected case can sit in a queue and then go to a specialist. Keep that in mind, because it matters as soon as you start putting numbers on them.

Put a number on it before you pitch it

Here is a synthetic example. It describes no client. It shows the arithmetic I would expect to see before anyone proposes a solution.

Suppose the workflow handles 30,000 cases a year, and 25% still need manual repair after extraction, mostly because information was missing at intake. Each repair takes 12 minutes on average.

That is 30,000 × 25% × 12 ÷ 60 = 1,500 hours a year of gross repair effort. At an assumed fully loaded cost of £35 an hour, it is £52,500 of gross effort value.

Synthetic example · Not client data or realised savings

30,000

Cases per year

25%

Still repaired by hand after extraction

12 min

Minutes per repair

£35/h

Assumed fully loaded cost per hour

Gross repair effort

1,500 h · £52,500

(30,000 × 25% × 12) ÷ 60 × £35/h

If 60% proves preventable (hypothetical)

900 h · £31,500

1,500 h × 60% × £35/h. Addressable capacity value a year, before the cost of the change

Capacity value is not a cash saving. Cash depends on what the client does with the hours: less overtime, a hire avoided or reduced agency spend.

That figure is not the opportunity. It is the size of the pattern. Not every repair can be prevented: some documents will always arrive late, and some cases genuinely need judgement.

Suppose a sample of repaired cases suggested, hypothetically, that 60% could be prevented by fixing intake. The addressable figure becomes 900 hours, or £31,500 of addressable capacity value a year, before the cost of making and running the change.

Capacity value is not a cash saving. The client only banks cash if those hours stop overtime, avoid a hire or reduce agency spend. If the team simply absorbs growth with the same headcount, that is genuine operating leverage, but it belongs in a different column of the business case. Say which column you mean before the client's finance team asks.

Start with the cheapest intervention that works

This is where AI firms are most tempted to overreach. You have found the problem and you have a team that builds AI, so the proposal seems to write itself. Resist that for a moment.

I work down a simple ladder and stop at the first rung that materially moves the number:

  • Stop: remove a request, check or report that nobody uses.
  • Simplify: ask for less, or ask for it where the information first arrives.
  • Standardise: one rule, one form and one definition of complete, across every team.
  • Workflow: route incomplete cases back before they reach expensive handling.
  • Automate: apply deterministic rules to deterministic checks.
  • AI: interpret genuinely varied documents or judgement-heavy cases, where it adds value.

In the synthetic example, a mandatory field and a completeness check at intake might remove most of the preventable repairs for very little money. That is a smaller invoice than an AI build. It is also the advice that gets you invited back, and it often clears the way for the AI work that genuinely needs doing.

A bounded ten-working-day review

You do not need a long discovery phase to do this properly. An illustrative shape for one workflow:

  • Working days 1–2: agree the boundary. One workflow, a named process owner on the client side and a shared definition of a repaired case. Request a month of completed cases, including the failures.
  • Working days 3–5: trace a sample of real cases from arrival to completion. Note where each one went back, waited, was re-keyed or went to a specialist, and why.
  • Working days 6–8: size each pattern from volume, affected share and minutes per case. Agree with the client's finance lead what counts as cash, cost avoidance or released capacity.
  • Working days 9–10: remove the overlaps, write up to three opportunity briefs and present a recommended next decision.

The output is not a report about the client's operation. It is a short set of decisions the client can take, each with the evidence behind it.

Three opportunity briefs, not a wish list

Each brief answers the same five questions. What is the problem? What is the evidence? What are the economics? Which intervention? What decision do we need next? Using the synthetic example:

Opportunity brief 01 · Synthetic example

Missing information at intake

Rework

Problem
A quarter of cases need manual repair after extraction, mostly for missing signatures and schedules.
Evidence
Repair reasons across a month of cases, checked against a traced sample.
Economics
1,500 gross hours a year; 900 hours and £31,500 of addressable capacity value if 60% proves preventable. Capacity, not cash.
Intervention
Simplify and standardise the intake request; a workflow rule returns incomplete cases before extraction.
Next decision
Approve a bounded intake pilot with the process owner, measured against the repair rate.

Opportunity brief 02 · Synthetic example

Exceptions waiting for one reviewer

Queues · Specialist bottlenecks

Problem
Unreadable or unusual cases wait for days before a senior reviewer picks them up.
Evidence
Timestamps comparing elapsed time with active handling time, and who resolves each exception type.
Economics
Sized as delay and specialist capacity, not repair hours. Cases already counted in the first brief are excluded.
Intervention
Standardise the common exception rules so others can resolve them; route the rest by type.
Next decision
Agree which exception types the wider team may resolve, and who signs off the rules.

Opportunity brief 03 · Synthetic example

Month-end reconciliation

Duplication

Problem
Extracted data is checked again against the finance system before anyone trusts it.
Evidence
A field-by-field comparison of what is keyed or checked twice, and how often the second check finds an error.
Economics
Sized from reconciliation hours, then reduced for errors the first brief would prevent at source.
Intervention
Stop the check if it rarely finds anything; otherwise automate a deterministic match. AI only if the documents genuinely vary.
Next decision
Ask finance whether an automated match with exception reporting would satisfy their control.

Do not add the three briefs together. A case repaired at intake may also have waited in the exception queue and been reconciled downstream. Count the effort once, against the cause that creates it, and show the client how you did it.

A total built from overlapping savings is the quickest way to lose a finance director's trust. A smaller number that survives challenge is worth far more than a large one that does not.

Make it repeatable across your client base

Done once, this is a good account conversation. Done consistently, it becomes how your firm grows. Every delivered build is followed by a structured look at the workflow around it: the same arithmetic, the same care about cash and capacity, and briefs your clients can approve or decline quickly.

The firms I see doing this well do not sell more AI. They sell better-evidenced next steps, and AI is one of the options on the ladder. Clients notice the difference, and they tend to remember who helped them make a sound decision rather than an expensive one.

For AI implementation firms

Make this review a repeatable way to grow existing accounts

AI-PEX is Leania's licensed delivery system for consultancies. It gives your consultants a structured way to find, cost and propose follow-on work with existing clients, with Leania guiding your first live engagement.

See how AI-PEX works

Want to try it on one account first? The free Client Opportunity Expansion Blueprint turns what you have observed into ranked opportunity briefs and a client-ready PDF. Open the free blueprint

← All insights