The Annual-Contract Problem: How to Sell Agents When Nobody Can Predict Usage

Table of Contents

• Why the SaaS Annual Contract Doesn't Survive Contact with Agents
• The Two Forecasts That Are Both Wrong
• Where the Risk Actually Lives
• Four Ways Vendors Are Bridging Commitment and Usage
• Annual Commit with a Drawdown Pool
• Committed Floor Plus Metered Overage
• Ramp Deals That Price the Unknown as a Curve
• True-Forward Instead of True-Up
• The Procurement Wrinkle Nobody Warns You About
• How to Choose Without Guessing
• Insights Most People Overlook
• References

Why the SaaS Annual Contract Doesn't Survive Contact with Agents

The annual SaaS contract was a beautiful piece of financial engineering. You bought 250 seats of a CRM for the year. Whether your reps logged in every day or never, the number on the invoice didn't move. Predictable for the buyer, predictable for the vendor, and the variance between "what we paid for" and "what we used" was somebody's problem to ignore at renewal. Seats are a stock. You count them once and they sit still.

Agent usage is a flow, and flows do not sit still. An agentic AI service, whether it resolves support tickets, reconciles invoices, runs outbound research, or files compliance reports, consumes resources every time it does work. The unit of value is a task or an outcome, not a login. And the number of tasks a business will hand to an agent over twelve months is not a knowable quantity at the moment you sign. It depends on the customer's own growth, on how much they trust the thing after ninety days, on whether their team finds three new workflows to point it at, and on how the agent's success rate changes as the underlying models improve.

That last factor is the quiet killer. With a SaaS seat, the product is roughly as useful in month eleven as it was in month one. With an agent, the product often gets dramatically more capable mid-contract, a model upgrade lands, the success rate on a hard task class jumps from 60% to 85%, and suddenly the customer wants to route ten times the volume through it. Good news for everyone, except the annual contract you signed assumed last quarter's volume. As a16z has noted in its writing on the shift from software margins to AI-service economics, the cost-to-serve in this category is variable in a way pure software never was, which means a fixed annual price is a bet on usage that one side will lose.

So you have a contract instrument designed for a static quantity being applied to a quantity that is not just variable but actively trending, often upward, often in ways neither party can see coming. The friction this produces is not a billing inconvenience. It shows up as stalled deals, sandbagged commitments, surprise overages that poison renewals, and finance teams on both sides who don't trust the number.

The Two Forecasts That Are Both Wrong

When a vendor and a buyer sit down to size an annual agent deal, they are each secretly running a forecast, and the two forecasts are constructed to be wrong in opposite directions.

The buyer forecasts low. Not out of malice, out of survival. A procurement-approved annual number becomes a budget line, and budget lines that get exceeded trigger pain: emergency approvals, awkward conversations with finance, the dreaded mid-year amendment. So the buyer commits to the volume they're confident they'll hit, which is usually their current volume minus a safety margin. They are pricing their own uncertainty as caution.

The vendor forecasts high, because their commit number drives their revenue recognition, their board metrics, and frankly their comp plan. A big annual commitment looks like a big customer. The salesperson is incentivized to anchor the commit at the customer's aspirational volume, "once you roll this out to all five regions", because that maximizes contract value today.

The gap between these two numbers is where the deal either dies or gets papered over with a discount. And here's the part that trips up teams coming from a SaaS background: neither forecast is more honest than the other. They're both rational responses to genuine uncertainty. The mistake is treating the negotiation as a fight over whose forecast is correct, when the real task is designing a contract that doesn't require either forecast to be correct. This is the same underlying problem explored from the buyer's side in discussions of enterprise procurement versus consumption pricing, and it's why the "just pick a number" approach keeps failing.

Where the Risk Actually Lives

It helps to be precise about what kind of risk you're actually trading, because "usage is unpredictable" smuggles together three different risks that want different contract treatments.

Volume risk is the obvious one: how many tasks. This is the variance everyone fixates on. It's also, ironically, the most forecastable of the three over a long enough window, because it tends to correlate with the customer's own business metrics, ticket volume, transaction count, headcount.

Cost-to-serve risk is the vendor's nightmare and is mostly invisible to the buyer. The price the vendor pays to run each task is tied to inference costs, model choice, and how many retries a task needs. A locked annual price exposes the vendor to margin compression if a customer's task mix shifts toward harder, more token-hungry work, a problem we treat directly in passing through volatile inference costs. Inference pricing has trended down over time, which sounds like it favors the vendor, but a multi-year fixed contract signed at today's costs can also strand the vendor on the wrong side of a repricing, which is the grandfather problem in miniature.

Mix risk is the sleeper. Two customers running the same number of tasks can have wildly different economics if one routes mostly easy, single-shot tasks and the other routes ambiguous, multi-system work that the agent has to attempt three times. Annual contracts almost never price mix, which means a customer can stay within their committed volume while quietly destroying the vendor's margin, or vice versa, paying for a fat annual commit while only sending trivial work.

Separating these three is the whole game. Most contract structures that work are really just mechanisms for assigning each of these risks to whichever party is best positioned to absorb it, rather than dumping all three onto a single fixed annual price.

Four Ways Vendors Are Bridging Commitment and Usage

There is no single right answer, but there are four patterns that show up repeatedly in real GaaS deals, each with a distinct philosophy about who eats the uncertainty.

Annual Commit with a Drawdown Pool

The customer commits an annual dollar amount up front and draws it down as they consume tasks, the way you'd burn through prepaid cloud credits. The commit gives the vendor the predictable revenue they need; the drawdown gives the customer the freedom to use the agent however and whenever they want without renegotiating.

The elegance here is that it decouples commitment from consumption timing. The buyer doesn't have to forecast their monthly cadence, only their annual envelope. This is the model that the broader credits and prepaid pools pattern is built around, and it has become something close to a default for usage-heavy AI products. The cloud hyperscalers normalized it years ago; AWS's committed-spend agreements trained an entire generation of finance teams to think in drawdown terms, which is half the reason buyers accept it now.

The catch: what happens to unused credits? If they expire, the buyer feels robbed and remembers it at renewal. If they roll over indefinitely, the vendor's revenue recognition gets messy and the "commitment" loses its teeth. Most workable versions land on partial rollover or a use-it-or-lose-it floor with a generous overage runway on top.

Committed Floor Plus Metered Overage

The customer commits to a baseline annual volume at a discounted rate and pays metered overage for anything above it. This is the hybrid model done right, and it's the most intuitive structure for buyers crossing over from SaaS because the floor feels like a subscription.

The reason it works under unpredictable usage is that it asks the buyer to forecast only the part they're confident about, the floor, and lets the unpredictable upside flow through as variable cost. The buyer sandbags the floor a little, which is fine, because the overage mechanism catches the rest. The vendor gets a predictable base and uncapped upside.

The failure mode is overage shock. If the per-unit overage rate is meaningfully higher than the committed rate (which vendors love, because overage is high-margin), a customer who blows through their floor gets a nasty invoice and a justified grievance. The fix is to keep overage pricing within shouting distance of committed pricing and, critically, to alert the customer before they cross the line, not after. Overage that arrives as a surprise is a trust tradeoff you usually lose.

Ramp Deals That Price the Unknown as a Curve

Instead of one annual number, the contract specifies an increasing commitment over the term: low in Q1 while the customer is still onboarding and trust is building, higher by Q4 once the agent is embedded. The total contract value is real and bookable, but the curve acknowledges that usage in month one is not usage in month twelve.

Ramps are underused in agent deals and they fit the category unusually well, because agent adoption genuinely does ramp, there's a deployment-and-trust period before volume scales, a dynamic explored in pricing pilots versus production deployments. A ramp contract turns that adoption curve from a forecasting problem into a contractual feature. It also gives the salesperson a clean story: you're not asking the customer to commit to year-twelve volume on day one; you're committing to a trajectory.

The risk is that the back end of the ramp is a forecast wearing a contract's clothes. If the customer's adoption stalls, the Q4 commitment they signed becomes the same overcommitment fight, just deferred. Ramps work when the ramp is grounded in a concrete rollout plan, region by region, team by team, rather than a hopeful upward line.

True-Forward Instead of True-Up

This one is subtle and it's where the smartest vendors are quietly moving. In a traditional true-up, you reconcile actual usage against the commit at period end and the customer pays for the difference, which means overages feel like penalties. A true-forward instead takes consistent overage as a signal to increase the committed baseline going forward, often at a better rate, rather than billing the gap punitively.

The psychological difference is enormous. A true-up says "you went over, pay up." A true-forward says "you're clearly getting more value than we sized for, let's right-size your commitment and reward you for the bigger number." One is a clawback; the other is an automatic land-and-expand motion that the customer experiences as a discount rather than a bill. Salesforce and other large vendors have used true-forward language in their enterprise agreements for years precisely because it converts overage friction into expansion revenue without the renewal ambush.

The Procurement Wrinkle Nobody Warns You About

Even if you design the perfect flexible structure, you run into a wall that has nothing to do with pricing logic: enterprise procurement is built to approve fixed numbers, and a consumption-variable contract is genuinely hard for it to process.

A procurement org's entire control apparatus, the purchase order, the budget approval, the spend authorization, assumes a known commitment. Hand them a contract whose total cost depends on usage they can't see yet, and you trigger a different, slower approval path, often with a hard cap requirement attached. This is why so many beautifully designed usage contracts end up with a not-to-exceed ceiling bolted on, effectively recreating the annual fixed problem from the other direction. Gartner's research on software and cloud contract negotiation has long flagged that procurement's tooling lags consumption models by years, and agents have widened that gap.

The practical move is to give procurement a number they can put on a PO, a committed floor or an annual envelope, while keeping the flexibility in the mechanism underneath. Procurement gets its fixed line item; the contract terms handle the variance. You're not fighting procurement's need for a number; you're satisfying it with the right number and absorbing the unpredictability in the structure around it. This tension is the heart of the enterprise procurement standoff, and pretending it's solvable with pure pricing cleverness is how good deals die in legal review.

How to Choose Without Guessing

If you're a vendor staring at this menu, the selection logic is less about your pricing preference and more about which of the three risks, volume, cost-to-serve, mix, dominates your situation.

If your cost-to-serve is stable and volume is the main variable, a committed floor with metered overage is the cleanest fit; you can price overage confidently because each marginal task costs you a known amount. If your cost-to-serve is volatile because task difficulty swings, lean toward a drawdown pool priced in your own units (resolved tickets, completed reconciliations) rather than raw tasks, so mix risk gets absorbed into the unit definition instead of your margin. If you're selling into a net-new use case where adoption is the real unknown, a ramp deal grounded in a concrete rollout plan beats any attempt to forecast steady-state volume.

And regardless of structure, instrument the relationship so that being wrong about the forecast is cheap and visible. Show the customer their consumption trajectory in real time. Alert before thresholds, not after. Build the true-forward conversation into the quarterly business review so expansion feels like a mutual right-sizing rather than a renewal ambush. The contract structures above are all just scaffolding; what actually determines whether unpredictable usage destroys the relationship is whether the customer ever gets surprised by a bill. The vendors winning this category have mostly concluded that the annual contract isn't dead, it just has to stop pretending it knows the future, and start being honest that it's a frame for a moving number rather than a fixed one.

Insights Most People Overlook

• The "unpredictable usage" framing is partly a measurement failure, not a forecasting failure. Most vendors can't forecast a customer's usage because they've never instrumented the customer's upstream driver, the ticket volume, transaction count, or headcount that actually generates agent tasks. Tie your forecast to the customer's own business metric instead of to historical agent volume, and "unpredictable" usage becomes surprisingly predictable. The variance was never in the agent; it was in the demand feeding it.

• Overage is the single most damaging line item in the entire GaaS pricing toolkit, and vendors love it precisely because it's invisible until it isn't. A high overage rate is high-margin revenue right up until the moment it lands on a customer's desk as a surprise, at which point it costs you the renewal and the reference. Treat overage rate as a relationship liability to be minimized, not a revenue lever to be maximized, the math that makes it attractive on a spreadsheet is the same math that makes it lethal at renewal.

• A model upgrade mid-contract can be a financial event, not just a product one. When your agent's success rate jumps because a better model shipped, customers want to route more volume, which is great, but a fixed annual contract converts that good news into either margin compression (you eat more inference) or an awkward repricing conversation. Smart vendors write a "capability uplift" clause that lets both the price and the committed volume re-baseline when success rates cross a threshold. Almost nobody does this yet.

• The true-forward versus true-up distinction is worth more than any headline rate. Two contracts with identical pricing but different reconciliation language will produce radically different renewal rates, because one frames overage as a penalty and the other frames it as earned expansion. The reconciliation mechanism is doing more emotional work than the price, and it's nearly free to change.

• Procurement's discomfort with consumption pricing is a feature you can sell against. When your competitor offers pure usage-based pricing, their deal stalls in the buyer's procurement queue for weeks. If you show up with a committed envelope that procurement can stamp plus the flexibility underneath, you can win on speed-to-close alone, even at a worse nominal rate. The annual commit isn't just a financing instrument; it's a procurement lubricant.

References

• The Economic Case for Generative AI and Foundation Models, Andreessen Horowitz (a16z)
• AWS Savings Plans: Committed-Spend Pricing, Amazon Web Services
• Cloud Strategy and Contract Negotiation Insights, Gartner