Pricing Pilots vs. Production Deployments: The GaaS Trap Hiding in Your Pilot Invoice

Table of Contents

• Why pilot and production are two different products
• The four pricing mistakes that kill the conversion
• How to price the pilot itself
• Designing the bridge to production economics
• What changes structurally at production scale
• A worked example: support agent, pilot to prod
• Insights Most People Overlook
• References

Why pilot and production are two different products

There's a persistent fiction in agent sales that a pilot is just "production, but smaller." It isn't. A pilot and a production deployment are two different products that happen to share a codebase, and pricing them as if they're the same thing is the single most common own-goal I see GaaS vendors commit.

The pilot's job is to retire risk. The buyer doesn't yet believe your agent resolves 40% of support tickets, or books appointments without hallucinating, or closes the books without a human double-checking every journal entry. The pilot exists to convert "maybe" into a number on a slide. Its currency is evidence, and its time horizon is six to twelve weeks. Nobody renews a pilot. They either kill it or scale it.

Production's job is to deliver durable value at a margin you can defend. Its currency is reliability and unit economics. Its time horizon is the multi-year contract. The buyer who signed the pilot to learn something now has to defend a budget line to a CFO who wants to know the cost per outcome three years out.

Those are different buyers with different fears, even when they share a name and an email address. Pilot-buyer is curious and forgiving. Production-buyer is accountable and skeptical. Price the same way to both and you'll over-serve one and under-serve the other.

This connects directly to a theme that runs through this whole pricing cluster: in GaaS, how you price reshapes the entire sales motion. (We dig into that specifically in the piece on how agent pricing changes the sales motion entirely.) The pilot-to-production seam is where that reshaping is most violent, because it's the one moment the buyer re-evaluates everything with real data in hand.

The four pricing mistakes that kill the conversion

I've watched enough of these deals stall to recognize the failure patterns. Four show up over and over.

The free pilot that anchors at zero. Founders love giving the pilot away "to remove friction." The problem is that free isn't a price, it's a reference point. A customer who paid nothing for sixty days of value has been trained that the value is worth nothing, and your production quote now reads as a tax on something they were getting for free. Anchoring research has been clear on this for decades; the first number a buyer sees disproportionately shapes every number after it. A free pilot anchors at zero and you spend the entire production negotiation climbing out of a hole you dug yourself.

The pilot priced at production unit economics. The opposite error. The vendor, terrified of margin erosion, charges full per-outcome rates during the pilot. But a pilot runs at terrible unit economics by design, low volume, heavy hand-holding, custom integration work, an account engineer practically living in the customer's Slack. Pricing the pilot to cover all that makes the pilot look absurdly expensive per unit, and the buyer extrapolates that ugly number into production. You've sabotaged the very comparison you needed to win.

No pre-agreed conversion price. The pilot succeeds, everyone's delighted, and then you start negotiating production pricing from scratch. Now you're negotiating from a position of having already proven you're valuable, which sounds like leverage but actually isn't, because the customer has also learned exactly how your sausage gets made and has a year of internal champions who will frame any increase as gouging. The price should be agreed before the pilot starts, contingent on hitting success criteria.

Success criteria that don't map to the pricing metric. A pilot measured on "did the team like it?" and a production contract priced on "resolved tickets" are measuring different things. When the pilot's success metric and the production billing metric diverge, the conversion conversation becomes a re-litigation of what the agent is even for. McKinsey's work on scaling AI from pilot to production repeatedly flags this measurement gap, the QuantumBlack analysis of why AI pilots stall keeps landing on the same root cause: the thing the pilot proved isn't the thing the business needs to buy.

How to price the pilot itself

So what do you actually charge for a pilot? My strong bias: never free, rarely cheap, always time-boxed.

Charge a fixed pilot fee, a flat, predictable number, rather than usage-based metering. Three reasons. First, a flat fee creates skin in the game; a customer who wrote a check assigns staff, opens systems, and shows up to the readouts. Free pilots get ghosted. Second, a flat fee removes the buyer's anxiety about runaway costs during exactly the phase when they're least able to predict usage. (The broader dynamic here, that metered billing breeds buyer anxiety, is worth its own treatment, and it gets one in the psychology of metered billing and buyer anxiety.) Third, a flat fee decouples pilot economics from production economics, so the ugly per-unit math of a low-volume pilot never poisons the production comparison.

Size the fee to be material but not a procurement event. The sweet spot is "expensable without a committee", large enough to command attention, small enough that the champion can sign it without dragging in legal and a six-week vendor review. For mid-market that's often $5K-$25K; for enterprise it might be $50K-$150K. The point isn't the revenue. The point is commitment and a clean anchor.

Crucially, make the pilot fee creditable against the first production invoice if they convert. This is the move that reconciles "charge for the pilot" with "remove friction to scale." The buyer hears: you're not paying twice, the pilot fee just becomes a down payment on the real thing. It turns the pilot fee from a cost into a deposit, and deposits convert.

Designing the bridge to production economics

The bridge is the whole game. Here's the structure I'd push every GaaS founder toward.

Agree the production price model and rough numbers in the pilot contract. Not a vague "we'll discuss commercial terms", actual ranges tied to actual outcomes. "If the agent resolves ≥35% of tier-1 tickets at ≥90% CSAT during the pilot, production pricing is $X per resolved ticket with a Y/month floor." The buyer signs up for the success criteria and the price they'll pay if those criteria are met, simultaneously. This is the difference between a pilot and a glorified free trial.

Tie the production metric to the value the pilot demonstrated. If the pilot proved hours saved, price production on hours saved or a clean proxy. If it proved resolved tickets, price on resolutions. The whole point of a value-based model, covered across this cluster in pieces like value-based pricing when the value is a replaced employee, is that the customer's spend tracks the value they receive. That alignment has to survive the pilot-to-prod transition or the model breaks at the worst possible moment.

Build in a ramp, not a cliff. Even with pre-agreed pricing, going from a $15K pilot to a $400K annual contract in one renewal is a budgetary heart attack. Stage it: a ramped first production year that starts near pilot economics and steps up as usage and proven value grow. This also hedges the real risk that production usage takes a quarter or two to reach the volumes everyone projected.

Decide who owns the integration cost. Pilots run on duct tape and goodwill. Production needs real integrations, SSO, audit logging, the works. Be explicit about whether that hardening is bundled into production pricing or billed as a one-time implementation fee, because surprise professional-services line items are a classic conversion killer.

a16z has written persuasively that the durable AI businesses will be the ones where pricing maps to delivered work rather than seats; their analysis of the new business models emerging around AI agents is a useful frame for thinking about what the production contract is really selling, output, not access.

What changes structurally at production scale

Beyond the headline number, several things genuinely change once you cross into production, and your pricing has to anticipate them.

Inference cost volatility becomes your problem, not a rounding error. At pilot volumes, model costs are noise. At production volumes, a model price change or a routing decision can swing your gross margin by ten points. This is why production contracts need margin protection baked in, pass-through clauses, floors, or routing logic, in a way pilots never bother with. (The mechanics of surviving volatile inference costs get a dedicated treatment in margin-safe pricing: passing through volatile inference costs.)

Reliability expectations harden into SLAs. A pilot tolerates the agent being wrong sometimes; everyone's learning. Production buyers want commitments, and commitments have a price. Higher reliability tiers, faster escalation, guaranteed uptime, these become pricing levers that didn't exist in the pilot.

The buying committee expands. The pilot champion could sign a $15K check. The production deal pulls in procurement, security review, FinOps, and often the line-of-business owner whose headcount the agent is implicitly affecting. Each new stakeholder brings a new pricing concern, predictability for procurement, auditability for security, cost-per-outcome for FinOps. Your production pricing page has to answer all of them at once.

Usage growth becomes automatic, and that's a double-edged sword. A successful production agent expands its own footprint; more workflows route to it without anyone signing a new order form. That's the land-and-expand dream, but it also means the customer's bill grows on autopilot, which breeds resentment if it's not capped or transparent. Production pricing needs guardrails, caps, alerts, prepaid pools, that a pilot never needed.

A worked example: support agent, pilot to prod

Make it concrete. A vendor sells an AI support agent on a per-resolution model, the Intercom Fin playbook, essentially.

Pilot. Flat $20K for an eight-week pilot, scoped to one product line and tier-1 tickets. Success criteria written into the contract: ≥40% deflection at ≥88% CSAT, measured on a defined ticket cohort. Production price pre-agreed: $0.99 per resolution with a $4K/month floor, ramped over the first two quarters. The $20K pilot fee credits against the first three production invoices.

What this buys the vendor. Real money on the table (committed customer, not a tire-kicker). A clean flat anchor that doesn't expose ugly low-volume per-resolution math. A pre-agreed production price, so a successful pilot converts on rails instead of reopening negotiation. And a credit mechanism that makes "pay for the pilot" feel like "pre-pay for production."

What this buys the customer. A capped, predictable pilot cost with no metering anxiety. Crystal-clear success criteria, so they know exactly what "it worked" means. A production price they've already blessed, so there's no renewal ambush. And a ramp that hedges the risk that production volume takes time to materialize.

The conversion conversation then isn't a negotiation. It's a status check: "We hit 43% deflection at 90% CSAT, that clears the bar, here's the production order form at the price we agreed, your $20K credits against the first invoices." That's how you make production a foregone conclusion.

Insights Most People Overlook

A paid pilot converts better than a free one, and it's not close. The reflex is that free removes friction and therefore lifts conversion. In practice the opposite holds: free pilots suffer from no internal champion accountability, get deprioritized the moment something urgent lands, and anchor the perceived value at zero. The pilot fee isn't revenue strategy, it's qualification. A customer who won't expense $15K to test something they claim could save them 40 hours a week was never going to buy. Charging filters out the tourists before they waste your engineering team's quarter.

The pilot's per-unit economics are a trap you set for yourself. Every founder's instinct is to show the customer the real per-outcome cost during the pilot "for transparency." But pilots run at structurally terrible unit economics, low volume, high touch, and showing that number trains the buyer to expect a price that production volumes would never justify. Quote the flat pilot fee. Quote the production per-unit price. Never let the buyer compute the pilot's per-unit price, because it's a lie about the steady state.

Success criteria are pricing terms in disguise. Teams treat pilot success criteria as a technical/product conversation and pricing as a separate commercial one. They're the same conversation. The metric you choose to define success is the metric you'll be forced to bill on, because that's the number the customer internalized as "the value." Choose a success metric you'd be happy to price against at scale, or you'll win the pilot and lose the margin.

The grandfather risk starts at the pilot, not the renewal. Whatever price you pre-agree in the pilot contract becomes the price the customer feels entitled to forever, even as model costs drop and your own economics shift. The pilot is where you quietly plant repricing rights (annual adjustment clauses, model-cost pass-through language) that you'll desperately want two years later. Founders who skip this to "keep the pilot contract simple" inherit a base of customers locked into yesterday's economics. This dovetails with the broader repricing problem the cluster examines elsewhere, the pilot is round one of it.

A failed pilot is a pricing data goldmine, if you instrument it. Most vendors treat a lost pilot as a binary loss. But a pilot that hit 31% deflection against a 40% bar tells you precisely where your value curve sits and what price the customer would have paid for that lower performance. The teams that win the next ten deals are the ones that price the pilot to generate that curve, capturing performance-versus-willingness-to-pay data, rather than just chasing a single yes/no.

References

• McKinsey QuantumBlack, Insights on scaling AI from pilot to production
• Andreessen Horowitz (a16z), Business models and pricing for AI agents