Before you lock your 2027 roadmap: ten checks to run on every AI idea before it gets a date.
5 tests · 10 checks · about 10 minutes per idea
How to use it
Run one copy per AI idea, in about 10 minutes, with your product lead and your engineering lead in the room. Every AI idea on your roadmap probably sounds reasonable on its own. This checklist tests whether an idea will carry real work for your users, and whether you know enough to give it a date.
Who
The person who owns the roadmap, plus one person each from engineering and design. Answer the checks together; disagreement on a check is useful information in itself.
When
Before an idea goes on the dated roadmap. It is most useful in planning season, when many ideas compete for the same engineering quarter.
How
Tick a box only when you can point to evidence or name a specific person. Where a check asks two things, tick it only when both are true. "Probably" counts as unticked.
Then
Place each test in one of the three self-assessment categories, and compare your ideas side by side.
The checklist
There are five tests with two checks each, giving ten in total. Write the idea in one line at the top of your copy before you start.
AI idea:
Test 1
Whose work does it carry?
An AI feature earns its place when it moves real work for a specific person. When nobody can name that person, the feature usually turns into a panel that people open once.
1
Can we name the role, and the one task this shortens for them?
Why it matters
AI built for "users" in general rarely serves anyone well. A named role and a named task give design, engineering and sales one person to build for, and make success measurable later.
What good looks like
The idea is written with a role and a task in it, such as "helps the support lead triage the morning queue", and today's rough time for that task is known.
What to check
Ask three people separately who this is for and which task it changes. Then complete the line "Today this takes ___; with the feature it takes ___", filling the first blank from observation.
Warning sign
The answer is "everyone", or the task is described as "insights", "visibility" or "productivity".
2
Does the output reach the next person, inside the tool where they already work?
Why it matters
Most work in B2B software passes between people. If the AI's output stops with the first person, or sits in a separate panel, the handoff stays manual and the value leaks away.
What good looks like
The next recipient is named, and the output appears at the step it informs, in a form they can use without re-work.
What to check
Follow one real piece of work from the first person to the last, and mark on a sketch of each screen where the AI's output appears.
Warning sign
The output ends as a summary someone copies into another tool, or "it will be in the assistant panel" is the answer for every use.
Test 2
What happens today without it?
The best AI opportunities sit where a person is quietly bridging a gap by hand. Find the bridge before you design the feature.
3
Have we watched the task, and found where someone bridges the gap by hand?
Why it matters
Teams often design from how they believe a task runs. Watching even a few people reveals the copying, re-typing, searching and chasing where AI pays for itself.
What good looks like
Notes or recordings from several real sessions in the last few months, with a specific step named where someone leaves the product or re-enters data.
What to check
Ask for the most recent session where someone on the team watched this task being done, and circle every manual bridge in it.
Warning sign
The workflow comes only from internal staff or sales calls, or no bridge can be found, which may mean the task already works well enough.
4
Do we know what that gap costs, and does it show up across many customers?
Why it matters
A cost turns an idea into a business case and gives you a baseline to measure against. A gap shared across accounts justifies product investment, while a gap in one large account can pull the roadmap towards one customer.
What good looks like
The cost is stated in a unit the business tracks, such as minutes per task, error rate, wait time or renewal risk, and the same gap appears in accounts of different sizes or segments.
What to check
Estimate the cost for one customer over one month, showing your working. Then search support tickets, call notes and customer-success records for the gap across accounts.
Warning sign
The cost is described only as "frustration" or "friction", or every mention traces back to one customer or one deal.
Test 3
What's the evidence?
Competitor launches and board questions create pressure. Customer evidence tells you where that pressure should go.
5
Have customers described this problem in their own words, before any competitor launched?
Why it matters
Customers talk about problems, while teams talk about solutions. A competitor's launch creates urgency, and urgency is easily mistaken for demand.
What good looks like
Verbatim quotes from calls, tickets or reviews describe the pain without suggesting AI, and the problem was on record before any competitor's announcement.
What to check
Find three verbatim customer quotes about this problem, and the earliest date it appears in your own records.
Warning sign
Every quote names the feature because the interview question did, or the idea entered the backlog the week a competitor announced theirs.
6
Can we get the data the AI needs, today?
Why it matters
An AI feature is only as good as what it can see. Missing or messy data can multiply the timeline.
What good looks like
The data exists, is accessible, and someone has checked its quality on real records.
What to check
Ask engineering to pull a sample of the data the feature would use.
Warning sign
The plan relies on data that customers would first have to supply or clean.
Test 4
Does it fit the product?
Every new AI feature competes for the same screen and the same user attention. Fit decides whether the product feels like one system.
7
Do we know where it lives, and does it follow our AI principles?
Why it matters
Placement tells users what a feature is for. AI features built separately also develop different personalities, and a mistake by one erodes trust in all of them.
What good looks like
The team agrees on placement and can explain it in one sentence, and a written set of AI principles covers what the AI does, what it never does and how it speaks.
What to check
Ask product and design separately where it lives, and compare the answers. Then hold the idea against your principles and note every conflict.
Warning sign
Placement is decided by which team happens to build it, there are no written principles, or this feature needs an exception on day one.
8
Is acting on its output faster than doing the task by hand, and can users correct it?
Why it matters
Reading, checking and correcting AI output all take time, and when that time exceeds what the feature saves, people quietly stop using it. People trust AI they can check, and their corrections show you how to improve it.
What good looks like
In a quick test, users act on the output faster than they finish the task by hand. Each suggestion shows its source, and editing or rejecting it takes one action.
What to check
Time three people doing the task with a prototype, then without it. Trace the path for a wrong suggestion, from spotting it to fixing it.
Warning sign
Users must verify every line before they can act, or the only way to handle a wrong answer is to ignore it.
Test 5
Can we size it and reverse it?
AI work rarely has a past project to anchor an estimate. Size the smallest proof first and know your way back.
9
Has engineering sized the smallest release, with the deferred items written down?
Why it matters
Without a sized first release, the full vision sets the timeline. Deferrals that live only in conversation creep back into scope.
What good looks like
The first release is described in a paragraph with an engineering estimate, plus a short list of what sits outside it, agreed by product and engineering.
What to check
Ask engineering for the estimate and its assumptions. Share the deferred list with sales and customer success, and confirm they accept it.
Warning sign
The only estimate on the table is for the complete feature, or sales has already promised a deferred item to a customer.
10
Have we agreed one success measure, and do we know what reversing it would cost?
Why it matters
A measure chosen after launch tends to flatter the feature. Customers also build habits, workflows and contracts around AI features, so the cost of reversal tells you how carefully to commit.
What good looks like
One measure is defined before launch, with a baseline from Test 2, such as time to resolve or handoffs completed. The team can name what a reversal would affect: customers, pricing, data and integrations.
What to check
Write the measure and today's baseline on the roadmap item. List who would notice if the feature disappeared, and what you would tell them.
Warning sign
The only measure is clicks, opens or usage, or the feature goes into pricing or contracts before it has proved itself.
Your self-assessment
When you've worked through the checks, place each of the five tests in one of three categories for each idea. Base the choice on the "What good looks like" and "Warning sign" lines, and on the evidence you could show a colleague today.
Working well
The business has a clear process or supporting evidence in place.
You can name the people, point to the evidence, and the team gives the same answers.
Needs attention
Some elements exist, but important gaps remain.
Some answers rest on evidence; others rest on belief, one customer or one meeting.
Investigate further
The business lacks enough information to understand the problem or its impact.
Nobody knows the answer yet, or the team disagrees and has nothing to settle it with.
Treat the result as an initial self-assessment. It reflects what your team knows and believes today. A definitive diagnosis would need a closer look at your product, your customers' work and your data, and this checklist does not claim to be a formal audit.
Test
Idea A
Idea B
Idea C
1 · Whose work does it carry?
2 · What happens today without it?
3 · What's the evidence?
4 · Does it fit the product?
5 · Can we size it and reverse it?
The gap that worries us most
Reading the pattern
Investigate further on Test 1 or 2. The user problem itself is still unclear. Watching the work comes before any other step.
Working well on Tests 1 and 2, with gaps in Tests 3 to 5. A real problem with an unfinished plan. Close the gaps before the idea gets a date.
Working well across all five. The idea is ready for a roadmap decision, and your team can explain why.
If two ideas carry the same person's work at the same step, check whether they should be one feature.
Pause the idea
Red flags that override any score
If any of these is true, pause the idea, however strong the rest of it looks. Each one usually means the team is responding to pressure and has not yet found a problem behind it.
The strongest argument for it is "a competitor just launched this."
It is described by its technology. For example, "add an agent" or "use an LLM for X", with no user or task in the sentence.
Its only success measure is usage. People can open a feature often while their work gets no faster.
Nobody can say what the AI should never do here. Without a boundary, the feature drifts and trust erodes the first time it oversteps.
It depends on data you don't have yet, and nobody owns getting it.
The first release is the whole vision. Nothing has been deferred, so nothing can be learnt before the big commitment.
What to do with what you find
Finding a gap is the first step. Most teams find several, and they rarely matter equally. The real work is deciding which gaps put the most engineering time, customer trust or revenue at risk, and which actions to take first. Your team can make that call with this checklist alone; start with the idea your roadmap is about to commit to.
If you'd like to talk it through, message me on LinkedIn with the word ROADMAP.
We'll discuss which of these areas may be worth examining in your product, look at how your team approaches AI decisions today, and agree on one useful next step.