The Bolt-On Bot
A GPT wrapper stapled to the page. Tell: it cannot touch your real data, permissions, approvals, or workflow.
The AI slop problem
Most AI shipped today is slop.
Measured, owned, and senior, or it is not AI. It is theater you rent forever.
AI slop is output that looks like work and costs like work but was never measured, never owned, and never touched by anyone senior. It is a demo that survives exactly one path. It is a prompt wrapped in a logo. It is a model that gets vaguer the harder you press it.
Slop passes the eye test in the meeting and fails the second question in production. You can spot it by what is missing: no baseline, no error rate, no named owner, no incident history, no proof that last week's change made the system better instead of louder.
The DPR line: if an AI system cannot name the number it moves, show the evals that guard it, and survive inside your real workflow, it is not transformation. It is a subscription to ambiguity.
Reference bar
Clean like Vercel. Serious like Harness. Direct like Devin.
The category does not need more neon dashboards. It needs senior builders who can sit beside revenue, operations, support, legal, and engineering, then ship one useful capability that keeps working after the sales call ends.
Seven slop types
The tell always shows up before the invoice.
Lazy vendors sell mystery because mystery protects margin. Senior teams sell measurement because measurement survives procurement, board questions, and production traffic.
The Dead Demo
Flawless on the call, absent from production. Tell: no live metric, no user count, no incident history.
The Rented Brain
You pay per call forever and own nothing. Tell: no weights, no code, no prompts, no exit path in the contract.
The Vanity Accuracy
They say “98% accurate” with no floor under it. Tell: no sample size, no comparison group, no confidence interval.
AI-First Theater
A rebrand with the word AI and no behavior change. Tell: the roadmap is slides, not commits.
The Transformation Retainer
A monthly agency wrapper around a generic model. Tell: deliverables are decks and “strategy,” never a moved number.
The Eval-Free Ship
A model in production with no test suite. Tell: nobody can say if last week's change made it better or worse.
The Slop Sniff-Test
Run this on any vendor. Run it on us first.
Founders do not need a PhD to detect fake AI. They need eight blunt questions and the patience to wait through the silence after each one.
- Ask for the baseline. What was the error rate, cycle time, cost, or conversion rate before AI, and what is it now?
- Ask who is senior and named. If the senior person only appears during sales, you are buying supervision theater.
- Send three off-path inputs. Change the wording, add a messy exception, include missing data, and watch what happens.
- Ask what happens when it is unsure. Serious systems refuse, route, or ask. Slop guesses with confidence.
- Ask what you own if you cancel tomorrow. Code, prompts, evals, fine-tune, docs, deployment path, and runbook should not disappear.
- Ask for the eval set. The answer should include the last five failures it caught before shipping.
- Ask how they know it improved last month. “The model is newer” is not an answer. A repeatable measurement is.
- Ask the failure mode they fear most. Be suspicious of anyone who names none. They have not operated it long enough.
A good vendor will welcome this test because it makes the work smaller, sharper, and safer. A slop vendor will call it “enterprise complexity” and try to move you back to the deck.
Their slop vs our discipline
The difference is not taste. It is ownership.
AI slop
DPR discipline
Ships a demo that survives one scripted path.
Ships against an eval set that covers your eight worst cases before go-live.
Calls it intelligent with no number attached.
States a baseline and the error rate, then reports the change the same way each time.
Hands you a junior team and a Slack channel.
Names the senior builder who owns the work and writes production code.
Guards a single prompt as the whole product.
Gives you the prompts, evals, code, and runbook so the capability is yours.
Answers every question at full confidence.
Designs refusal, escalation, and human review for the cases that need it.
Points at a stale benchmark and calls it proof.
Measures on your traffic and shows the failures caught before release.
Prices automation, delivers hidden manual rewrites.
Tells you exactly where a human is in the loop and why.
Locks capability behind a forever invoice.
Builds so that if you fire us, the working system keeps working.
The Owned-AI Method
Number. Baseline. Build inside. Evals. Handover. Proof.
The antidote is not a bigger model. The antidote is a smaller promise, made in public, with a scoreboard attached. We begin by choosing one number worth moving: hours saved, tickets resolved, quote time reduced, error rate lowered, lead quality improved. If a number does not matter enough to measure, it does not matter enough for AI.
Then we establish the baseline before writing the solution. We build inside your systems, not beside them, because the hard part is rarely the model call. The hard part is permissions, messy edge cases, human approval, audit trails, rollback, latency, and the boring path that keeps the work alive on Tuesday morning.
Evals come before victory laps. We write the cases your users will actually send, including ugly inputs and adversarial examples. We track regressions. We log failure modes. We build the handover while we build the system, so your team receives the code, the prompts, the runbook, the measurement method, and the right to fire us without losing the capability.
- Number: one business metric, chosen before the build.
- Baseline: the current state measured from real work, not a guess.
- Build inside: integrated with the workflow, data, permissions, and people who use it.
- Evals: tests for quality, refusals, drift, and the failure cases that cost money.
- Handover: repo, docs, runbook, measurement, and operating ownership.
- Proof: the same number reported after launch, with caveats in plain sight.
The closing test
If nobody owns it, nobody built it.
Run the sniff test on your current vendor. Then run it on us. If we cannot show you a baseline, an error rate, an eval set, and a named senior who owns the work, walk away. That is the standard, and it is the whole pitch.