Almost nobody loses a deployment to an architecture problem. They lose it to a data owner who never says no but never grants access, a manager whose team the system automates, a sponsor who changes jobs in month five, and a set of users who quietly kept the spreadsheet. This lab gives you six real-shaped people at one customer. First read the room — classify each one from what they say, not their title. Then handle the three situations that actually end projects, while a coalition meter tracks who is still with you.
Meridian Logistics has bought a shipment-exception triage system. Six people were in the kickoff. Each said one sentence that tells you what they actually are. Classify them — titles will mislead you at least twice.
1/4
act
0
score
—
coalition
The five roles, and why titles lie
One person can hold two roles. A senior title can hold none. What identifies someone is not where they sit on the chart but what they are accountable for and what they stand to lose.
Role
How to recognise them
What they need from you
Economic buyer
Talks in outcomes and money, not features. Often has not been in a working session since kickoff.
A number, twice a quarter, in their language. Never a demo.
Champion
Defends the project when you are not in the room. Their reputation is now attached to it.
Ammunition — evidence they can forward, and no surprises. Protect them publicly.
End user
Asks specific questions about their actual Tuesday. Sceptical in a detailed way.
To be consulted before the design is fixed, and to not look stupid in front of colleagues.
Blocker
Raises process objections. Rarely says no outright; says “that will need to go through…”
To be engaged early, on their terms, in writing. Late engagement is what makes them adversarial.
Data gatekeeper
Owns the system you need. Cautious, specific, and has been burned before.
A narrow written request and evidence you will not misuse what they give you.
most underestimated
gatekeeper
most fragile
single champion
quietly decides
the users
costs most
late escalation
Why Palantir handed forward-deployed staff a book on improv
The often-repeated detail is that Palantir gave new forward-deployed hires Keith Johnstone’s Impro — a book about theatre, status and accepting offers. It reads as a joke until you have sat in a customer meeting where the real subject was who would be seen to have been right.
01
Status is being negotiated constantly
Who defers to whom, who interrupts, who gets to reframe the problem. In a customer’s room you have no formal status at all — only the status you are granted — and a great deal of what happens next depends on reading that accurately in the first ten minutes.
02
“Yes, and” beats “no, but”
Blocking someone’s idea in front of their colleagues costs you more than the idea was worth. Accept the offer, then extend it: “yes — and if we do that, we will need X, so the trade is Y.” Same information, no bruise.
03
Make your partner look good
The improv principle and the deployment principle are identical. Your champion presenting your results as their win is the best possible outcome, not a slight. People fight to keep projects that make them look effective.
04
Read what is actually happening in the room
The objection stated is often not the objection held. “I am concerned about data quality” can mean exactly that, or it can mean “my team produces that data and you are about to audit it in public”. Responding to the wrong one loses you the person.
The three situations, and the honest answers
# 1. the gatekeeper who never says no
instinct escalate to the sponsor
result access in ~a week, an enemy for ~18 months, and every
future request routed through the slowest legitimate path
better ask what their process is · put a NARROW request in writing
(tables, columns, read-only, retention) · offer masked or
synthetic data while approval runs · give them a way to say yes
# 2. the manager whose team you are automating
instinct reassure — "nobody's job is going away"
result you cannot promise that, it breaks, and you lose the room
better treat the resistance as information · find the version that is
good for them (exceptions not volume · capacity for work they
never got to · their team OWNS the system) · give them a win
# 3. the champion resigns in month five
instinct carry on and hope the successor inherits the enthusiasm
result a new manager reviewing commitments they did not make
better while they are still there: get value-to-date IN WRITING in
front of the economic buyer · ask them who else cares ·
then build the second champion you should have had by month four
Notice the shape all three share: the instinct is to treat the person as an obstacle, and the better move is to treat them as someone with a rational position you have not yet understood.
The picture: what each act moves
The lab’s scoring, read from its own code, on the best path. Two gauges decide the ending. The coalition is the average of four people’s meters. The judgement score is Act 1’s six classifications plus each act’s points. Each act moves only the people it touches, so the act aimed at Tomas cannot make up for having alienated Ruth. Alan, the CTO, has no meter at all.
Worked example: two runs through Meridian, meter by meter
Run A follows instinct: four of six people classified right, then escalate, reassure, carry on. Run B classifies all six and takes the best move in each act. Meters out of 100. The coalition is their rounded average, and the score is out of 27.
Run
Step
Ruth
Tomas
Dana
Marcus
Coalition
Score
A
After Act 1 (4 of 6 right)
45
40
55
70
53
12
A
Escalate over Ruth
15
40
50
75
45
10
A
“Nobody’s job is going away”
15
15
50
75
39
7
A
Carry on, meet the successor in January
15
15
40
75
36
5 (19%)
B
After Act 1 (6 of 6 right)
45
40
55
70
53
18
B
Ask Ruth what passes
70
40
55
70
59
21
B
Refuse the promise, offer the 30%
70
65
60
75
68
24
B
Value summary to Dana
70
65
80
75
73
27 (100%)
Run A is Cancelled in the January review (coalition 36). Run B is Renewed and expanded. Look at Marcus in run A: he is the only meter that went up, to 75, because escalating was his idea. The champion was happy the whole way to cancellation. The judgement score matters on its own too. Take run B’s three act choices but misread three of the six people: the coalition is still 73, but the score is 18 of 27 (67%), under the 70% bar, and the outcome drops to Survived, with a harder second year.
Cost and trade-offs of each move
Here Time is what the move costs you this week, and Space is what it leaves behind, as the lab’s meter changes.
Case
Time
Space
Why
Escalate over the gatekeeper
Fastest route to this one grant
Ruth −30. Every later request runs at the board’s pace
Borrowed speed. You pay it back on every request for the rest of the engagement.
Ask the gatekeeper what passes
A 20-minute meeting and a resubmission
Ruth +25, an ally
Her process is legitimate. Treating it that way is what makes her fast next time.
Masked data now, full request in parallel
Unblocks today
Ruth +15, Marcus +10
Pragmatic, but you still do not know what the board wants, so the real request can bounce again.
Reassure the threatened lead
One sentence
Tomas −25, and a promise you do not control
Costs nothing to say and breaks the first time headcount moves.
Refuse the promise, offer the interesting 30%
A harder conversation
Tomas +25, Marcus +5, Dana +5
The only option that reaches three meters. Honesty about what you cannot promise is what makes the rest believable.
Value summary before the champion leaves
One document and one email
Dana +20
The record outlasts Marcus and reaches the one person who can say yes alone.
Ship features for the successor
Weeks of work
Tomas −10
Destabilises a working system just when its record needs to look boring.
The cheap moves are cheap only this week. Everything that costs one sentence now costs a relationship later. The moves that cost a meeting or a document are the ones the meters reward.
Common mistakes with stakeholders
Reading a happy champion as a healthy deployment
In run A above, Marcus ends higher than he started while every person who decides the outcome falls. Your champion is the person most likely to tell you things are fine. Track the gatekeeper, the threatened manager and the buyer separately.
Counting logins, not whether the old process died
Analysts can log in every morning and still do the real work in the spreadsheet. Adoption means the old queue or file stopped being updated. Ask to see it.
Taking a yes in the room as a commitment
People nod in meetings and act on their own priorities afterwards. Follow every agreement with a short written note: who does what, by when. It turns a mood into a commitment someone can be held to.
Letting bad news land first in a big meeting
A missed target your champion hears about at the same time as their boss costs them credibility they spent on you. Tell them privately first, with the fix, so they can present it themselves.
Treating the kickoff list as the stakeholder list
Procurement, legal, HR and the owner of a downstream system often attend no kickoff and turn up in month three with a veto. Ask early who else has to sign off before this goes live, and who would be surprised by it.
Where this shows up in real-world work
Enterprise sales and renewals
Sales teams formalise the same map. MEDDIC, a common qualification framework, makes “Economic buyer” and “Champion” two of its six checks, and account plans list each stakeholder’s role, stance and influence. The lab’s meters are that map, updated after every move.
Internal platform migrations
The five roles appear inside your own company. The owner of the old system is the gatekeeper. The team whose tool gets replaced is the threatened manager. The VP who funded the migration and then moved on is the champion who resigned.
Automation projects
Any system that takes over part of a team’s work runs Act 3. The questions to settle before launch are which cases stay with people, who owns the system afterwards, and how the team’s expertise gets credited. HR or employee representatives may also need to be consulted.
Behavioural interviews
“Tell me about a time a stakeholder blocked you.” Strong answers describe the person’s rational position before the tactic, show a narrow, written request or a refused false promise, and end with the relationship intact, not just the access granted.
Check yourself
You classify all six correctly, then take each act’s second-best good move: masked data, the never-automate boundary, and a meeting with Dana. What is the outcome?
Renewed, but only just. The meters end at Ruth 60, Tomas 60, Dana 70 and Marcus 75. That averages 66.25, which rounds to 66, against a bar of 65, and the score is 24 of 27 (89%). The second-best moves each reach fewer meters than the best ones, so there is very little margin left.
Same three moves, but you misread two of the six people. Does the outcome change?
Yes, it drops to Survived. The coalition is still 66, but the score falls to 18 of 27 (67%), under the 70% that Renewed needs. Misreading people costs points even when it does not cost a relationship in the lab, because in real life you would have spent your time on the wrong person.
In week 10, procurement says the renewal needs three competing quotes before it can be signed. Which role is that, and what do you do?
A blocker: it can stop the renewal without owning the outcome. Engage early, on its terms and in writing. Ask what the process requires and when, and give Dana the value evidence she needs to justify the choice. Arguing that the process should not apply is the escalation mistake from Act 2.
Ruth has sent your access request back twice with questions. Is she blocking you?
No. Sending a request back with questions is a process doing its job, and she is the gatekeeper, accountable for the data. The signal to act on is that the request was too broad. Ask her what passes, then resubmit a narrow, read-only request naming the tables and the retention period.
Common questions
Who are the key stakeholders in an enterprise deployment?
Five roles, and one person can hold more than one. The economic buyer controls budget and can say yes alone. The champion spends their own credibility internally. The end users decide whether it is used. The blocker can stop it without owning it — usually security, legal or procurement. The data gatekeeper controls access to the systems you need, and is the most consistently underestimated.
How do you handle a data gatekeeper who will not grant access?
Escalating is the instinct and it is usually wrong: it makes an enemy of someone you need for the rest of the engagement, and they can comply slowly and legitimately for months. Ask what their approval process actually is, send a narrow written request naming exact tables, columns and read-only scope, and offer to work with masked or synthetic data while approval proceeds.
What do you do when your champion leaves?
Move while they are still employed: get a written summary of value delivered in front of the economic buyer, and ask the champion directly who else in the organisation cares about this outcome. Their successor inherits a commitment they did not make, so the project needs a second relationship in a different reporting line — ideally built in month four rather than in the week they resign.
How do you work with a manager whose team you are automating?
Their resistance is rational; treat it as information. Do not promise that nobody’s job changes — you do not control that, and the promise destroys your credibility when it breaks. Find the version of the outcome that is good for them: their team handling exceptions rather than routine volume, taking on work they never had capacity for, or owning the system themselves.
Why do deployments fail politically rather than technically?
Because the technical problem is usually solvable and the organisational one often goes unaddressed. A system nobody adopted, an access request that took five months, a sponsor who moved teams, or a manager quietly encouraging their staff to keep using the old process will each end a deployment regardless of how well the software works.