Show Me the Roadmap You Cut: Five Tests for Buying From a Software House

TL;DR

• “Five developers from Monday” no longer sells, because your own team has the same AI tools those developers would use.
• Five tests ask for evidence that already exists: the backlog item the vendor argued against, a signed delivery clause, the incident count on their last system, their own worst project, and a handover package your engineer can rehearse with.
• A software house can commit to delivery by a date, not to your business result, so sign fixed-price releases with agreed remedies and refuse outcome clauses the vendor defines and measures itself.
• In finance and insurance the exit test is a compliance item, because DORA has required termination, exit and transition terms in contracts behind critical functions since January 2025.
• When continuity, a balance sheet that can pay a penalty and patterns from other clients do not matter to you, one senior engineer with good AI tools is the better buy.

“We have five developers available from Monday.”

That sentence used to close deals. Now your own team has the same AI tools those five developers would use, so every vendor you meet will tell you the same thing instead: judgment, ownership, industry knowledge, production experience.

A claim that costs nothing to make tells you nothing about the company making it, so what you need is a document, a person or a clause that already exists and cannot be produced on the spot.

I am on the vendor side of this table, so I have an interest in how that meeting ends, and the tests below work against my own company when they should.

Why the meeting changed

At a technology conference in summer 2026, a panel on AI in large companies got the question of where AI had disappointed. One panelist, a director responsible for the P&L of a key business in one of the largest technology groups in Poland, answered that his teams had become more productive while his vendors’ invoices stayed the same, and he came back to it more than once, in public, with some of those vendors in the room.

He is not alone, because Bain, the consulting firm, wrote in its Technology Report (September 2026) that clients now demand as much as 20% efficiency improvements at renewal and award new work assuming at least 30% lower delivery cost. Vendors concede less: Forrester, the analyst firm, quoted in October 2026 K. Krithivasan, chief executive of TCS, the largest Indian IT services company by revenue, saying the productivity gain passed on to clients is in most places in the 10 to 15% range. Your next negotiation sits between those numbers, and five tests tell you who is still selling hours under a new name.

Test 1: Judgment. Show me the roadmap you cut.

If you are renewing, ask for the item from your own backlog that the vendor argued against building because it would not move a business number, and was right about, then ask for the email, ticket or meeting note where they said it at the time. The item must have been paid work for the vendor, because a proposal to “phase” a feature is a reprioritization, not a cut.

If the vendor is new, share a bounded part of the backlog and ask them to cut it in the proposal. Adding missing security or migration work is fair, but a vendor who only adds is pricing hours, and a list of roles and day rates sells you capacity you already have, which in 2026 is like buying a calculator from your accountant: you already own one, and you were paying him for the answers.

They will say “NDA”, and they may be right, so accept a client-approved extract or a walkthrough, because all you need to know is whether the decision was recorded.

A bad answer sounds like: “We always challenge the scope in discovery.” You asked for a document.

Test 2: Accountability. Show me the signed clause.

We can’t guarantee your business result, but we can commit to delivery: a working system, in production, by an agreed date, with real money at stake if it slips. Ask for a redacted delivery clause from a signed contract, with the counterparty’s signature block visible and the name blacked out, because a template from the vendor’s standard agreement proves only that they own a word processor. Have counsel check the remedy under your governing law, and write your own obligations next to it, environments, approvals, test data and acceptance criteria, because a date without those is a promise neither side can keep.

If the vendor offers outcome-based pricing instead, Bain (August 2026) lists three conditions for it, observable, uniquely attributable and contractible, and says software engineering productivity remains difficult to measure consistently. Ask who defined the metric, who measures it and what else moves it, because if the vendor’s dashboard alone decides the penalty, you carry the risk.

A bad answer sounds like: “We are open to outcome-based pricing.” Ask for the signed one.

Test 3: Production. Ask the reference about the first six months.

Skip the call with the sponsor who signed. Ask the reference client to put its IT lead on the call, after shortlisting and before award, and ask how many production incidents the vendor’s last system caused in the first six months, who handled them at night, and whether the vendor was on the incident bridge. A regulated client will not hand out its on-call engineer’s number, but its IT lead knows the count, and when the client picks the person you get a witness rather than a reference.

A bad answer sounds like: “Our delivery lead will walk you through the production history.”

Test 4: Pattern knowledge. Ask what went wrong, and check it with someone the vendor did not pick.

Ask the vendor for a project that went wrong and what they changed afterwards, and listen for their own mistake, dated, and a process change still in use, because “the client changed the scope” is a story about you, not about them. Then put the same question to the references they gave you, since anything longer than a few months has had a bad week, and add one former client from your own network if you can, because references chosen by the vendor only show you what the vendor chose to show.

A bad answer sounds like: “We have never had a failed project.” After years of work, that is untrue or a sign they only take easy jobs.

Test 5: Exit. Show me the handover package from a finished contract.

In finance or insurance this test is a compliance item: DORA, the EU regulation on digital operational resilience in application since January 2025, requires contracts for critical or important functions to cover termination, exit and transition, with a documented and tested exit plan for each (FinTech Global, May 2026). A clause only promises a handover, so ask for a client-approved handover package from a contract that ended: code, pipeline definitions, environment configuration, secrets handling and runbooks. Then give it to one of your engineers for a day in a safe environment to build, deploy and change one small thing without calling the vendor.

A bad answer sounds like: “Everything is in the client’s repository.” Ask what “everything” contained last time.

If you have thirty minutes

Run Test 1 on your own backlog and Test 4 on the references you already have, because neither can hide behind an NDA.

Why not one senior engineer with good AI tools?

Often you should. A company earns its margin on three things one person cannot give you: continuity when that person is ill or leaves, a balance sheet that can pay the remedy from Test 2, and patterns from several teams at different clients at the same time. Check all three rather than assume them, and if none matters to you, the engineer is the better buy.

What to sign, what to refuse

Sign:

• a fixed-price delivery of the next release, with a date, a remedy and change control
• a smaller, more senior team, priced as such
• ownership of code, prompts, documentation and AI output on your work, disclosure of the AI tools used, and no training on your data (the last three are among the contract areas a lawyer named in InformationWeek, April 2026)
• at least one reference you found yourself, before signature

Refuse:

• five developers from Monday at last year’s rate
• an outcome clause where the vendor defines and reports the outcome
• a model recommendation from a vendor that will not say which provider partner badges it carries, ours included

Using this on us

Run these tests on us as hard as on anyone, and if we cannot show you one of the five artifacts, we will say so in the first meeting and write the gap into the proposal. The vendors worth shortlisting are the ones who can put all five on the table, and one question sits behind all five, so ask it about us too: if hours were free tomorrow, would you still call?

If you found this interesting, you might want to read my previous articles: Are You Subsidizing Your Vendor’s AI Investments? and Forward Deployed Engineering: New Packaging, Same Bill?

Sources

• Bain & Company (Megha Chawla, Mark Brinda, Sushant Khandelwal, Sandeep Nayak) - “Tech Services: Cracking the Code on AI-led Growth”, Technology Report 2026, 29 September 2026

• Forrester (Ted Schadler) - “For Service Providers, AI Giveth And AI Taketh Away”, 16 October 2026

• Bain & Company (Eric Maltiel, Josh Sandberg) - “AI Pricing: A Reality Check on Effort, Usage, and Outcomes”, 10 August 2026

• EIOPA - “Digital Operational Resilience Act (DORA)”, Regulation (EU) 2022/2554, in application since 17 January 2025

• FinTech Global - “DORA compliance: closing the gap between contract and exit”, 6 May 2026

• InformationWeek (Stephanie Overby) - “Outsourcing contracts weren’t built for AI. CIOs are renegotiating now”, 28 April 2026