What’s the Real Cost of AI-Powered Deal Screening?

Updated August 2026 · 7 min read · Deal Room Intelligence Series

What opaque pricing tells you

When prices are hidden, price is set by perceived value to the buyer rather than by cost or by competitive reference. Practically, this means the same product is sold to a mid-market advisory firm and a bulge-bracket bank at figures that may differ by an order of magnitude, and neither knows.

Three consequences you can act on:

• Your first quote is an opening position, not a price. It reflects your perceived budget — inferred from your firm's name, headcount and deal sizes — far more than it reflects delivery cost.
• You cannot benchmark, and neither can your procurement team. Any “market rate” someone cites is anecdote. We checked; the published data does not exist.
• Disclosing your deal volume early is expensive. It is the primary input to the number you will be quoted.

The five costs, only one of which appears in the quote

1. Licence

Per-seat, per-deal, per-document, per-page, or an annual platform fee. The model matters more than the number, because it determines who on your team can use it. Per-seat pricing quietly caps adoption: the analyst who would benefit most is the one nobody wants to buy a seat for, so the tool ends up used by two people and evaluated as if it were used by the team.

2. Volume overage

The number quoted assumes a document or page volume. Deal flow is lumpy — three transactions in a quarter, then nothing. Ask specifically what happens above the included volume, and whether unused capacity carries forward. This is where annual contracts most often disappoint.

3. Verification labour

The permanent cost nobody quotes. Someone has to check escalated findings against source text, and that is not a transitional expense that disappears once the team is trained. Grounded commercial legal AI has been measured hallucinating between 17% and 33% of the time in an adjacent task^[1] — an error rate nobody in this category has eliminated. Any cost model without a standing verification line is wrong.

This cost varies enormously by tool, and it is the largest hidden differentiator between two products with similar licence fees. Output that quotes its source sentence is verified in seconds; output that writes fluent unattributed prose is verified by re-reading the document, which is the work you paid to avoid.

4. Integration and configuration

Getting documents in and results out. Thresholds set to your risk appetite. Categories calibrated so the tool stops flagging clauses your firm has waved through for a decade. Mostly one-off, but real, and typically borne by your most expensive people.

5. Security and procurement review

Your infosec assessment, DPA negotiation, the training-prohibition term. For a firm handling deal documents this is not optional, and for a first AI vendor it is often the longest part of the process.

A worked example — put your own numbers in

Assume a firm running 12 transactions a year, averaging 1,200 documents each, with a reviewer rate of $300/hour and a throughput of 12 documents per hour.

• Full manual coverage: 1,200 ÷ 12 = 100 hours per deal × $300 = $30,000 per deal.
• With screening, reading the top-ranked 25% in full plus a 10% sample of the rest: 300 + 90 = 390 documents ÷ 12 = 32.5 hours × $300 = $9,750 per deal.
• Difference: $20,250 per deal, or $243,000 across twelve.

Now subtract the costs the sales conversation omits. Assume verification of escalated findings adds 6 hours per deal ($1,800), and one-off integration and configuration costs 60 hours ($18,000) in year one.

• Year-one net, before licence: $243,000 − (12 × $1,800) − $18,000 = $203,400.
• That is the number the licence fee has to fit inside — and it is the number you should walk into the negotiation holding, rather than asking what the product costs.

Change the reviewer rate, the deal count, or the share you read in full, and this moves a lot. Set the read-in-full share to 60% and most of the benefit evaporates. That sensitivity is the actual finding here, and it is why a single industry-wide ROI percentage is meaningless.

The cost that dominates and never gets modelled

Everything above is workflow arithmetic. It is not where the real money is.

The largest financial effect of screening is deals abandoned earlier. If a blocking-category issue surfaces in week one instead of week four, you save the entire remainder of the diligence workstream, the associated adviser fees, and the opportunity cost of a team that is now free. On a single dead deal that can exceed a year of licence cost.

We cannot put a number on it, and neither can anyone else — attributing an abandoned deal to the moment of discovery is not cleanly possible, which is precisely why it is missing from every vendor ROI model despite being the biggest term. Model it as a range, and note that it only exists if screening runs before the expensive workstream, not alongside it.

Questions that move the price

1. What is the per-unit rate above the included volume, and does unused volume carry forward? Lumpy deal flow punishes annual commitments.
2. What does a single deal cost, standalone? Even if you intend to buy annually, the answer tells you the real unit economics and gives you a floor.
3. Is pricing per seat, and can we add read-only reviewers free? Per-seat pricing that taxes reviewers suppresses exactly the usage that produces value.
4. What is included in year one — configuration, threshold calibration, training? These are negotiable and frequently priced separately after signature.
5. Is there a pilot on our own documents, at a real price? A free demo on vendor-chosen documents tells you nothing. A paid pilot on a closed deal where you know every material issue tells you everything.

Where cheap becomes expensive

The lowest licence fee frequently carries the highest total cost, for reasons that only appear after purchase:

• Output without source quotes — the licence saving is consumed several times over by verification labour, permanently.
• No coverage reporting — you cannot tell an examined-and-clean document from an unprocessed one, so either you re-check manually or you carry an unmeasured gap.
• Silent truncation on long documents — the credit agreement is the document that matters most and the one most likely to be quietly cut short.
• Non-reproducible output — if the same file scores differently on two runs, you cannot defend a decision, and the cost lands as risk rather than as hours.

Bottom line

You cannot benchmark a market that hides its prices, so stop trying. Instead, build your own number: hours by activity, your real reviewer rate, an honest verification line, and a range for early abandonment. That figure is your ceiling, and it is specific to you.

Then negotiate against your ceiling rather than against their quote — because their quote was calculated from what they think you can afford, and yours was calculated from what the work is actually worth.

See our pricing, in public →