If you can write the if-then, the agent is a tax

Most of what teams call "agent spend" is AI doing formula work.

Clay launched Claygent in September 2023. By June 2025 it had passed 1 billion runs. After that, the same public write-up put the cumulative total at 5 billion. Clay's own growth team then said the quiet part out loud: when you can write the rule down completely, use a formula. Formulas are free and deterministic. Agents spend credits to reason.

If you can write the if-then, the agent is a tax.

What is formula work dressed up as an agent?

Formula work is any task where every case fits an if-then you could write on a whiteboard.

Employee count over 100. Title contains "operations." Best email from a waterfall. Contact already in the CRM. Merge three provider columns into one clean field. That is Signals hygiene. It is not judgment. It does not need a model that plans, acts, and adjusts.

Clay's own guide on agent use cases for GTM draws the line the same way. Formulas return the same output every time at zero credit cost. Agents buy answers a rule cannot cover. When you point an agent at a million-row rule, you are paying for theater.

I see founders do this because "agent" sounds like progress. The dashboard looks smart. The credit burn looks like investment. Then the review meeting asks why Convert still has a dirty list.

When does an agent earn the credits?

When the edge cases would take a hundred rules.

Messy titles. "Head of Growth" is not a keyword match for marketing or sales. Account sizing from a website that never prints headcount. Deal postmortems that read Gong transcripts and Salesforce history for the real lost reason. Account health that mixes credit burn, Slack tone, and call notes into a churn score. Persona labels from titles typed fifty ways. Free-text "how did you hear about us" that no picklist will ever capture.

That is judgment. That is where Clay's internal examples live: ABM research for tier-one accounts, competitive digests, pre-sales postmortems, post-sales health scores. Same spine as AI as operator in GTM. The agent is the operator on the messy row. The formula is the rail under the clean row.

If covering the cases would take a hundred rules, you want judgment. If covering the cases takes one rule, you want a formula.

How do you stop agent spend from eating Signals?

Run a credit autopsy once a week. Same hour as your signal review if you already have one.

1. Pull the top agent columns by credit burn.
2. Ask for each: could I write the if-then completely?
3. If yes, rebuild it as a formula. Kill the agent column.
4. If no, tighten the prompt to one job. One output schema. Test on ten rows before you burn the table.
5. Cap credits per agent. Clay's own builder guidance says the same thing: clear job, guardrails, version history, watch usage.

Do not "explore what's possible" on the full TAM. That is how you pay for runs of noise while your own ICP still has duplicate contacts.

Clay without the spam for founders is the same law from the send side. Clean inputs. Tight jobs. No vanity volume. Map of work before GTM agents is the map. Formulas and agents are two different tools on that map, not two brand names for the same spend.

What does this mean for Signals vs Convert?

Signals owns the score and the rails. Convert owns the note and the meeting path.

If Signals burns credits classifying "VP Marketing" with an agent instead of a formula, Convert inherits a slow, expensive list. If Signals uses agents only for messy titles, account fit, and the postmortem fields Convert actually reads, the list gets sharper without the tax.

Would you rather spend credits on judgment that changes who Convert calls, or on if-then work a free formula already solved?

Ehrenberg-Bass and the LinkedIn B2B Institute still put about 95% of B2B buyers out of market. Burning agent credits to re-label the 95% with prettier fields is not Signals. It is a tax on the 5%.

Adapt or fail. Write the if-then. Pay the agent only when you cannot.

Start Signals, Convert, Grow

FAQ

Is every Claygent a tax?

No. Claygent earns the spend on judgment work: messy titles, account research, deal postmortems, health scores, free-text classification. The tax is using an agent where a formula would return the same answer every time for free.

How do I know if my rule is complete enough for a formula?

If you can list the cases without "it depends" or "read the site and decide," it is a formula. If a junior hire would still need to interpret, it is an agent.

Does this mean I should cut agent budget?

Cut wasted agent budget. Keep the judgment budget. Clay's own cost framing starts the same place: ask whether the task needs AI at all before you pick a model.

Where does this sit relative to buying agents vs building rails?

Buy the agent, build the rails is the stack choice. This post is the credit choice inside the stack. Rails include formulas. Agents sit on top of rails, not instead of them.

How does this fit Signals, Convert, Grow?

Signals scores windows and keeps data cheap and clean. Convert uses the judged fields in the note. Grow scales the loop only after the if-then work is off the agent bill.