Audit and Iterate
Outcome: a monthly audit that attributes spend, measures what each job produced, and changes at least one thing on evidence rather than opinion.
- Surface
- MCP server
- Level
- Advanced
- Uses
check_credits- Credits
- 0
- Prerequisite
- A month of running lessons 04–06
Attribute the spend
Three buckets, and the split is usually not what people expect.
| Bucket | Typical share |
|---|---|
| Central scheduled jobs | 60–80% |
| Rep ad-hoc usage | 10–25% |
| Experiments and one-offs | 5–15% |
Most rollouts assume rep usage is the cost problem and it almost never is. A rep doing twenty briefs a month costs a fraction of one ungated weekly job. When the number is uncomfortable, look at the schedules first.
Separate workspaces per team make this attribution trivial, and it is the one genuinely good reason to split them.
Measure what each job produced
Cost alone tells you nothing. Pair it with output.
| Job | Cost per month | Produced | Cost per useful row |
|---|---|---|---|
| ICP scoring | ~1,000 | 60 qualified accounts | ~17 |
| Inbound triage | ~1,300 | 200 leads classified | ~6.5 |
| Signal digest | ~1,400 | 30 actionable signals | ~47 |
Then the harder question: what happened to the output? Sixty qualified accounts that nobody worked cost the same as sixty that produced meetings, and only one of those is worth repeating.
The four audit questions
Which job costs most per useful row?
And is that justified by what a useful row from that job is worth?
Which output is not being used?
A digest nobody reads, a scored list nobody works, a triage nobody looks at. Retire it or fix its distribution — those are the only two options.
Which prompt overran?
Find the run, find the missing cap, fix the shared skill.
What did reps ask for that does not exist?
The requests are your backlog, and they are better evidence than a roadmap.
Retire things
A rollout accumulates jobs and prompts, and nobody volunteers to remove any.
Retire:
- Any scheduled job whose output nobody has acted on in a month
- Any shared prompt nobody has run in a quarter
- Any signal type in the digest that has produced no action
- Any enrichment column in a central job that nothing reads
That last one is worth checking specifically. Central jobs accumulate fields “in case they are useful”, and each is a per-row charge on every run forever.
Adjust guardrails from evidence
| Evidence | Adjustment |
|---|---|
| Nobody hit the per-person budget | Raise it — the cap is producing caution, not saving money |
| One prompt caused most overruns | Fix the prompt, not the policy |
| Reps stopped using it after week two | A usability problem, not a discipline problem |
| A job’s cost grew month on month | A cadence or seen/unseen filter problem |
| Digest read rate fell | Routing problem — signals need named owners |
The pattern: most guardrail problems are design problems. Tightening a budget in response to a design flaw produces a team that uses the tool less and a flaw that remains.
The monthly review
Thirty minutes, five questions, one named owner:
- What did we spend, split three ways?
- What did each central job produce, and what happened to it?
- What are we retiring this month?
- Which prompt needs fixing?
- What did reps ask for?
Write the answers down. The value is in the series, not in any single month.
Do this now
Pull the month’s consumption
Split into the three buckets.
Compute cost per useful row per job
Ask what happened to each job’s output
Honestly.
Identify one thing to retire
Find the biggest overrun and fix its prompt
Adjust one guardrail on evidence
Schedule the monthly review
With a named owner and a slot.
Check your work
- Spend is attributed to jobs, reps and experiments
- Every central job has a cost per useful row
- You know what happened to each job’s output
- At least one thing was retired
- One guardrail changed because of evidence
Where this breaks
Auditing cost without auditing outcome produces the wrong decision every time. The cheapest job is not the best one — a digest costing 1,400 credits that produces three meetings is worth far more than a triage job costing 300 that produces a classification nobody reads. Always pair the spend with what it caused, and be honest when the answer is “nothing”.
Course complete
You now have a standardized rollout, a per-person budget derived from real usage, guardrails living in the prompts rather than in a policy document, three central jobs producing scored accounts, triaged leads and a weekly signal digest, and a monthly audit that changes something on evidence.
Where to go next:
| You want | Course |
|---|---|
| The rep-side workflows in detail | AI for Sales Reps |
| Agentic builds on the same surface | AI-Powered GTM |
| Every MCP tool, clustered by job | SyncGTM MCP Mini Course |
| The signal plays as full courses | Signals & ABM |
Reference for this lesson: Credits, check_credits, Workspaces, MCP prompting guide.