Agents that own business outcomes
Postlet is an applied research lab. Most agents repeat a fixed task. Ours own a KPI: they form hypotheses, design and launch the experiments that settle them, and recursively improve until the number moves.
Pipeline this quarter
$4.2M
18% ahead of target pace
Experiments running
6
Across 3 platforms
Proven this month
3
Budget moved to winners
| Experiment | Platform | Status | Lift | Confidence |
|---|---|---|---|---|
| Short copy vs long, EMEA prospecting | Meta | Winning | +12% CTR | 97% |
| Founders vs HR leaders audience | Exploring | +4% | 61% | |
| Brand bid cap, two matched regions | Measuring | Day 9 of 28 | — | |
| One market held out vs control | Not incremental | −1% | 95% | |
| Lead form vs website, North America | Meta | Winning | +31% SQL rate | 96% |
| Weekend delivery off, 4 campaigns | Queued | — | — |
Sample data
Discovered
A CRM conversion feed has been silent for 9 days. Fix drafted.
Approval needed
Shift 20% of EMEA budget to the winning audience.
Learned
Short copy wins on Meta prospecting but not on LinkedIn. Next tests narrowed.
The method
Research, not a workflow
Outcome owned
Qualified pipeline from paid
+20% this quarter
May move up to 15% of budget a week
Guardrail: cost per opportunity under $4k
Own
Start from the outcome
Each agent owns one number, such as qualified pipeline, cost per demo or resolution rate. It connects to the systems where that number is measured and learns how the business works before it changes anything.
Experiment design
Founder audiences reach pipeline at a lower cost
Sized: 42k impressions per arm, 18 days
Explore
Turn hypotheses into experiments
It forms hypotheses from the data and designs controlled tests to settle them: variants, segment splits and market holdouts, each sized to give a reliable answer.
Read against control
+14% pipeline per dollar
Corrected for 3 tests run at once
Measure
Measure what caused the change
Results are read against a control, with intervals and corrections for running many tests at once, so a lucky week is never mistaken for a win.
Budget follows evidence
Exploit
Scale what works, keep exploring
Budget and effort move toward what is proven, with your approval, while a share keeps testing what has not been tried yet.
Self-improvement
An agent that gets better at getting better
A workflow does the same quality of work on day 300 as on day one. This loop closes on itself instead.
Every experiment is kept, with its context
The design, the result, the interval and the state of the account when it ran. New hypotheses are posed against what has already been settled, so the same question is never paid for twice and a weak idea is ruled out before it costs anything.
The chooser is scored, not only the answer
Deciding what to test next is itself a decision with a cost. The policy that ranks candidate experiments is graded on what its own picks returned and re-weighted, so the agent's judgement about where to look improves, not just its conclusions.
Hill climbing, with a share held back for the unknown
Most of the budget goes to the best configuration found so far. A fixed share keeps probing far away from it, because an agent that only exploits converges on the best version of the wrong idea.
Every conclusion has an expiry
Results are re-derived as new data arrives. When an effect stops holding, the finding is withdrawn and whatever was built on it is unwound, before spend keeps following a result that has decayed.
Improvement per experiment settled
Cumulative lift in the owned outcome, against experiments settled
- Experiments settled
- 412
- Conclusions live
- 87
- Ruled out without a test
- 1,104
- Withdrawn as stale
- 23
Sample data
All of it inside limits you set
Each agent is given the number it owns, how much it may move and the guardrails it may never cross. Every change it proposes waits for approval, is logged with before and after values, and can be reverted in one click.
The research engine
Discover what no one thought to test
Most of what moves an outcome never reaches a dashboard, because nobody thought to look for it.
Each challenged by three independent reviewers. Sample findings.
Daily
Every open finding re-verified
Reviewed
By three independent reviewers
Approval
Required for every change
Performance marketing
Marketing, run as a continuous experiment
Paid media settles a test in days, so it is where this architecture has the most history behind it: LinkedIn, Meta and Google Ads, read on CRM pipeline rather than on clicks.
Top of funnel
Grow qualified pipeline, not just clicks
The agent owns pipeline from paid media. It tests audiences, offers and budget splits across LinkedIn, Meta and Google Ads, reads every result against pipeline in HubSpot or Salesforce, and moves budget toward what creates qualified opportunities.
What it owns
- Qualified pipeline and cost per opportunity
- The budget split across platforms, regions and audiences
- Results read on CRM pipeline, not platform-reported conversions
Qualified pipeline from paid · by week
Experiments started in week 5
Budget split, moved toward pipeline
Sample data
Spend efficiency
Find spend that is not working, and prove it
Every day the engine re-checks every campaign, ad set and ad, and a review step discards weak findings before they reach you. Where a saving is uncertain, it designs a holdout: pause a market, compare it with a matched control, and keep the money only if results hold.
What it finds
- Budget increases that bought no extra results
- Audiences too small for the budget behind them
- Tracking that stopped, or counts the wrong event
Budget raise bought no extra results
HighMost recorded conversions are logins
HighAudience too small for its daily budget
MediumHoldout test · 21 days
Pause 2 markets, compare with a matched control
Ranked by money at stake. Sample findings.
Creative
Copy and creative that keep learning
The engine learns which wording works in your own ads by comparing ads that share an audience and budget. It writes new versions to test beside the ones that trail their siblings, and Creative Studio produces fresh image and video.
What it tests
- Hooks, length and message, measured in your account
- Every variant checked for new claims and brand before it runs
- Tests sized in impressions and days, read with confidence
Now running
Stop managing spreadsheets and vendors. Hire, pay and manage in one place.
0.81% CTR
Variant A
Hire, pay and manage in one place. No more spreadsheets.
0.91% CTR · +12%
Illustrative example
Beyond marketing
Every team has a number to own
The architecture holds wherever outcomes are measurable and decisions repeat. These are the teams we are building agents for next.
- - Customer support
- - Finance and accounting
- - Procurement
- - HR
- - GTM
Customer support
Resolution, not ticket volume
Agents that own resolution rate and cost per ticket. They test routing, answers and escalation paths, and keep what resolves more issues the first time.
Routing test · billing tickets
Winner at 95% probability · rolled out
Illustrative example
Finance and accounting
A faster close with fewer exceptions
Agents that own the speed and accuracy of the close. They work through reconciliation and coding, and route every exception to a person for review.
- Bank reconciliationDone
- Invoice coding, 1,240 linesDone
- AccrualsIn progress
3 exceptions routed to a reviewer
Illustrative example
Procurement
Spend that stays on contract
Agents that own savings and compliance. They compare quotes and terms, flag spend outside agreed contracts and track what each negotiation returned.
Quotes · laptop refresh
| Vendor A | $42.0k | Net 30 |
| Vendor B | $38.4k | On contract |
| Vendor C | $45.1k | Net 60 |
$12k off-contract spend flagged this month
Illustrative example
HR
Hiring that keeps pace
Agents that own time to hire and quality of hire. They test sourcing channels and outreach, and measure which ones produce people who are hired and stay.
Hires per 100 candidates, by channel
Effort shifted toward referrals
Illustrative example
Go-to-market
Pipeline from every motion
Agents that own pipeline across outbound, partner and product-led motions. They test segments, messages and sequences, and shift effort toward what converts.
Outbound sequence test
Problem-first
6.1%
reply rate · 14 meetings
Feature-first
3.8%
reply rate · 8 meetings
Pipeline created +18%
Illustrative example
Return on investment
What changes when an agent owns the number
Experiments that never stop
Tests run continuously instead of once a quarter, each with its decision rule set in advance.
Resources follow evidence
Spend and effort move toward what is proven against a control, not toward whatever a platform chose to report.
Problems caught in days
Broken measurement, runaway costs and stalled delivery surface within days, not at the next review.
Measured on outcomes
Results are read on pipeline and business KPIs from your own systems, not on clicks and impressions.
Returns that compound
Each settled experiment narrows the next one, so the rate of improvement rises instead of flattening out.
Full control
Guardrails, approval on every change and a complete audit trail with one-click revert.