Two rungs of the build spectrum live here. AI-Enabled Application — AI as a capability inside software: understanding documents, finding grounded answers, drafting, classifying, recommending. The AI works; the workflow around it decides. Agentic Workflow — AI that takes bounded action: picking the next step, using the business's tools, coordinating the work — with people in the loop where judgment matters, and every action recorded. Both are built in fixed scopes with published price bands, on the same software-factory basis as everything else: AI produces the raw build, humans hold the gates that make it safe.
Clear priorities. Practical next steps. No pressure to build.
Choose by what the AI should be allowed to do, not by the technology.
Rung 07 · AI-enabled application
Document understanding, semantic search over a knowledge base, summaries with sources attached, classification and recommendations. The AI works inside the workflow; consequential output gets human review. $15,000–$40,000.
Rung 08 · Agentic workflow
An agent that validates, enriches, routes, drafts, reconciles, and escalates — with approval gates on consequential actions, permissioned tools, an audit trail, and a kill switch. $20,000–$55,000.
The difference
If a step needs a judgment call you can't write as a rule, that's agent territory. Much of what's sold as "agentic" in 2026 is really a workflow with a language model doing one step — we'll tell you when that's all you need (rung 03, from $2,000).
Not a sales pitch
Rules are cheaper, more reliable, and easier to trust. We use AI where understanding is required and agents where workflows are genuinely variable — and we'll say so.
Published investment
$15,000–$40,000
AI doing real work inside a controlled app — document understanding, grounded answers with sources, classification — with guardrails and a test set gating every change.
See the full pricing ladderPublished investment
$20,000–$55,000
AI that takes bounded action with people in the loop: permissioned tools, approval gates, audit trails, a kill switch.
See the full pricing ladderGuardrails come standard
The difference between a demo and a system the business can run on is not the model. It's what surrounds it.
The model sees only what it may — records, contracts, and prompts stay inside the boundaries you define.
Output is tied to sources you can check, over smooth talk you can't.
Consequential actions — a customer messaged, a record changed, money moved — require a person.
Every action recorded, so "why did the system do that?" always has an answer.
A fixed scenario set is re-run before any change ships, so improvements can't quietly break what worked.
The system can be stopped and handed to a human in one action. Someone is always accountable.
The first question is never "which model?" — it's "what is the AI allowed to do, and who decides?"
Map the workflow, the information it touches, and where a person must stay in the loop.
Define what the AI can see, what it can change, and what requires approval.
Assemble the capability with the guardrails, not after them.
Fix the scenario set that gates every future change.
Go live with observation, escalation, and a kill switch in place.
Refine on evidence — what the system did well, what it missed, what it should never have done.
AI builds earn their place when the workflow is defined, the volume is real, and someone is accountable for the outcome.
Good fit
Document understanding and classification, with human review on the odd cases.
Good fit
An agent that gathers, checks, flags anomalies, and drafts the exception list.
Good fit
Bounded autonomy with approval gates where actions have consequences.
Not a fit yet
Automation makes broken processes fail faster. Rung 03 or a review comes first.
The boundaries are worth making explicit before anything is built.
FAQ
Model and API fees are third-party costs — paid by the client, or handled near at-cost if you prefer. Published prices never include them.
FAQ
No. Data boundaries are part of the build — the model sees only what the workflow requires, and we document it.
FAQ
Automation follows rules you can write down. An agent decides the next step when rules aren't enough — which is exactly why it needs approval gates and audit.
FAQ
Often not. If a rule or an existing tool solves it, we'll tell you — that's rung 03 territory at a fraction of the cost.
FAQ
Fixed scopes with published bands, exactly like the other rungs. The review — or a conversation — names the specific scope before anything is quoted.
Show us where AI should act
We'll help you decide whether the right answer is a rule, an AI capability, an agent — or nothing at all.
Book a free review