What a Back-Office AI Diagnostic Should Hand You Before You Buy an Install

Most companies with 200 or more employees have already tried AI somewhere. A team bought seats for a chat tool. Someone built an automation that works when its author is in the office. Operations still runs on spreadsheets, inbox rules and one person who knows how the vendor invoices actually get approved.

Buying more tools seldom closes that gap. What closes it is one back-office workflow rebuilt so it runs on its own, with a named person accountable for it. The diagnostic is where that work gets scoped. If the diagnostic is vague, the install will be too.

Below is what a paid diagnostic should hand you before you commit to an install.

Why most AI projects stall before they reach the back office

Pilots tend to stall for the same few reasons. The scope is a theme (“use AI in finance”), not a workflow. Nobody on the client side owns the result. The data and system access needed to build it arrive late or never. And there is no written definition of done, so the project ends when the budget does.

None of these are model problems. They are scoping and ownership problems, and a diagnostic exists to settle them before money goes into a build.

The five things a diagnostic should hand you

By the end of a diagnostic, you should have five things in hand:

1. A ranked list of candidate workflows with the current steps, systems, handoffs and volume for each one. Volume comes from your records, not from estimates in a slide.
2. One recommended workflow for the first install, with the reason it was picked over the others.
3. A named owner on your side, the person who will approve the build, sign acceptance and run the workflow afterward.
4. An access and data list naming every system, login, sample file and approval the build needs, with a date each item is due.
5. A draft acceptance sheet listing the specific, testable conditions under which the workflow counts as live.

If any of the five is missing, you are not ready to price an install yet, and nobody else is either.

One workflow, not a roadmap

A 40-page AI strategy is easy to produce and hard to act on. A first install should cover one workflow from trigger to finished record: an invoice arrives and ends coded and routed for approval, or a new hire is accepted and ends provisioned in every system.

Picking one workflow forces the hard decisions early: which system is the source of truth, who approves exceptions, and what happens when the agent is unsure. It also keeps the cost and timeline small enough to judge on results.

At The Bench, an install covers one workflow. Further workflows are scoped separately.

Who owns it on day one

Every agent we build has a name, and every agent has a named human owner on the client team. The owner is not a sponsor who signs off once. The owner reviews exceptions, approves changes and is the person the runbook is written for.

If the diagnostic cannot name that person, the workflow is not ready. Without an owner, the agent runs until the first exception and then gets switched off.

What “done” means, in writing

The acceptance sheet is the most useful page of the diagnostic. It turns “the automation works” into conditions both sides can test. For example:

• The workflow processes a defined sample set end to end.
• Exceptions route to the named owner with the reason attached.
• The runbook is delivered, and the owner has walked through it.
• User acceptance testing (UAT) is complete and signed.

Write down the conditions before the build starts. Then “done” is a checklist, not a debate.

How pricing should work after the diagnostic

Once the scope, owner, access list and acceptance sheet exist, a fixed price is realistic. Here is how The Bench prices it:

• Diagnostic: $7,500 to $12,000, credited toward the install.
• Install: $35,000 to $75,000 fixed, for one workflow.
• Run: $5,000 to $12,000 per month for monitoring, exceptions and changes after go-live.

The install fee is at risk. Half is due at kickoff. The other half is due when the workflow is live against the signed acceptance sheet. That condition depends on the client supplying the data, the access and a named owner on the dates in the access list.

We do not promise ROI, hours saved or headcount outcomes. We promise a working workflow measured against the conditions you signed.

A short test before you sign anything

Before you buy a diagnostic from anyone, including us, ask:

• Will I leave with one named workflow, or a list of ideas?
• Who on my team will you name as the owner, and what will they do each week?
• What access do you need, and by what date?
• What does the acceptance sheet look like? Can I see a blank one?
• What part of the fee depends on the workflow going live?

If the answers are specific, the diagnostic is worth paying for. If they are not, keep the money.

The Bench builds working plumbing, not strategy decks. To see whether your back office is a fit, start a conversation. You can also read how we work or browse more dispatches.