The agent is cheap. Who finishes the work?

“At that price, we can afford to try it.”

Suppose you're comparing two proposals to check supplier documents. One has a small monthly software bill. The other costs more and claims it will leave less checking for your team.

You could approve the cheaper one before lunch. The amount is almost too small to argue about.

Then your operations manager asks who will handle the cases it cannot finish.

Good question. Especially if that person is the same person whose time the proposal says you're buying back.

There are companies where the software budget looks excellent while someone spends Friday afternoon opening every result, checking the original attachment and deciding whether it can be used. The invoice for the tool arrives once a month. The employee's contribution disappears into the ordinary working day.

You feel the cost later. A customer still waiting for an answer. A manager doing the checking after everyone else has left. The next useful project pushed into October.

I would want that work visible before signing either proposal.

Put the reviewer in the calculation

Take a deliberately simple example. These are assumptions, not client results or a quotation.

A team handles 1,000 cases a month, taking 12 minutes per case. That's 200 hours.

The proposed system prepares the check. A person still spends three minutes reviewing every case: 50 hours a month. One hundred cases need another ten minutes each to resolve a missing or conflicting document. Those ten minutes come on top of the review already counted.

Another 16 hours and 40 minutes.

So the team would still spend 66 hours and 40 minutes on the work. The potential release is 133 hours and 20 minutes, before any additional time spent maintaining the system.

That could be a useful project. It could give someone a Wednesday morning to deal with supplier problems before those problems reach a customer.

It also leaves rather more human work than “the agent handles 1,000 cases” suggests.

At an assumed loaded labour cost of €40 an hour, those released hours represent about €5,333 of monthly capacity value. Put that number on a slide and it starts looking like money available to spend.

Your bank balance has not changed yet.

If the same employees receive the same salary, the immediate benefit is time they can use differently. A cash saving needs something specific to happen: a paid overtime shift removed, for example, or an external service you can actually stop buying.

Where the benefit comes from extra business, count the contribution left after serving it. An additional order brings costs with it. And the same hour cannot simultaneously support that order and justify a labour saving elsewhere in the spreadsheet.

There is nothing disappointing about capacity. Having good people available for work you keep postponing is a serious reason to invest. Give that benefit an honest name and a destination.

Make checking easier before buying more intelligence

McKinsey's August 2026 analysis puts human oversight at 70 to 75 percent of variable costs in its banking customer-service example, compared with 20 to 25 percent for tokens. Those are sector-specific estimates. Your supplier checks need their own measurements. They do explain why a cheaper model bill can miss the largest expense.

Watch someone review a case. Are they judging an exception, or hunting through attachments to find what the system used? Put the relevant passage beside the proposed result, with its date and source. A reviewer who can inspect the evidence directly may finish faster without giving up an approval that matters.

Some checks need no language model. Compare a date against a deadline with ordinary software. Use a model where interpreting the supplier's explanation helps. Test a cheaper model on representative cases before routing work to it, including the cases it must escalate. Count extra checking and retries in that comparison.

A case marked complete can still come back

Before comparing the two prices, ask both suppliers to work against the same definition of finished.

For this example, the required documents have been checked against the agreed requirements. Missing information is recorded. The outcome is saved where the next person needs it, with the evidence supporting it. Anything requiring approval has reached the authorised person.

“A summary was generated” leaves several of those jobs open.

A test should also expose the awkward cases. The attachment with an old date. Two documents that disagree. A supplier whose trading name differs from the name on the record. Cases like these help reveal whether a person gets a clear exception or a confident answer that takes longer to untangle.

Measure the effort through to acceptance. Follow cases that get reopened. Count failed attempts in the cost even when they never become finished work.

And look at consequences separately. A missing field that delays an internal update and a mistake that authorises an unsafe action deserve different treatment. A low average handling cost cannot make an unacceptable failure acceptable.

This is where the more expensive proposal can earn its price. If it reliably reduces the work left for qualified people, the total operating cost may be lower. It still has to demonstrate that. A larger invoice proves nothing by itself.

Can the same capability earn its keep elsewhere?

McKinsey also points to volume and reuse as ways to spread fixed costs. That deserves a concrete question in the proposal: which part of this build could another workflow actually use?

A document-reading service might serve both supplier onboarding and renewals. Permissions, monitoring and an evaluation set may also be reusable. The approval rules still need checking for each use. Copying an agent into another department is not evidence that it can safely do the work there.

Volume changes the calculation. A recurring fixed fee divided across a busy month's accepted cases may look reasonable; the same fee across a quiet month can be expensive. Show both. Include support effort when the second workflow arrives rather than treating reuse as free.

There may be a new offer here too. If checking becomes affordable enough to serve smaller customers, test whether those customers want the service and what they will pay. The relevant gain is the contribution from delivered work, after the extra service costs. A larger addressable market in a presentation is still a hypothesis.

The first month is a poor forecast

Suppose software, usage and support cost €1,500 a month, with a €20,000 setup. Again, illustrative figures only.

Subtracting €1,500 from the capacity value produces an attractive-looking monthly balance. Dividing the setup cost by that balance gives you a payback figure with impressive precision and a missing premise: whether the business can realise the value assigned to those hours.

I would leave that payback claim unsigned until the destination of the time is clear.

Then check the months that make the proposal uncomfortable. Volume falls while the fixed fee continues. A supplier changes its document format. The person supervising exceptions takes leave. Someone needs to update the rules and check that the update worked.

Those events belong in the operating plan. Ask who handles them and what is included in the price. If your own team provides the support, put their time in the calculation too.

The useful number is the total cost of getting acceptable work finished over a realistic period. Keep setup visible, show ongoing costs separately, and compare against the current process at the same volume and quality standard.

You can then make a decision you can explain without borrowing the vendor's confidence.

Every quarter, reopen the calculation with actual volume, accepted cases and review effort. Check whether a model or supplier change improved the total cost, and whether the recovered capacity reached its intended use. Keep an owner for that review. An agent that still runs can nevertheless have stopped being worth its cost.

For an industrial business or B2B distributor, start with one recurring workload. Bring last month's volume, the time people actually spend checking it, and the exceptions that keep returning. Add the work you would give those people if they had room.

That is enough to start a serious commercial conversation about whether the project deserves a closer assessment.

Request a Strategy Session.