Automate the boring part before you buy the intelligent part
Half the workflows companies want an LLM for do not need one. Deterministic automation is cheaper, faster and does not hallucinate.
There is a reflex right now to reach for a language model the moment a process looks tedious. Often the tedium is structural, not cognitive — a form that has to be re-typed into a second system, a report that has to be assembled from three exports every Monday.
The rule we use
If a competent junior could write the rule down in one page, it is automation, not AI. Rules are deterministic, auditable, free to run and never have a bad day.
Reach for a model when the input is genuinely unstructured — free text, scanned documents, images, speech — or when the branching is too wide to enumerate.
The hybrid is usually right
Most systems we ship are mostly deterministic with a model in one or two spots:
- A workflow engine handles the routing, retries and integrations.
- A model handles the extraction from the messy PDF in the middle.
- A rule validates the model's output before anything is written.
That architecture is cheaper to run than an end-to-end LLM pipeline by an order of magnitude, and when it breaks, you can see exactly which step broke.
Where to start
List the ten most repeated tasks in one department. Mark each one rule, judgement or both. The rules are your first sprint — they will fund the judgement work that follows.
The Reciprocal Solutions
Automation Team
We design, build and operate AI agents, automations and custom LLM systems — and stay on the hook for how they behave in production.
Got a workflow that should be running itself?
Bring us the process, the constraints and the mess. We will tell you honestly whether AI is the answer — and what it would take to ship it.