Bring your own AI: the planning tool that works with the LLM you already pay for
Somewhere in the last two years, every project management tool grew an AI button, and next to the button, a price: per seat, per month, on top of the licence you already pay. The AI behind the button is usually the same handful of frontier models you already subscribe to directly. You are paying twice for the same intelligence, and the second payment buys you a worse interface to it.
We think the structure of that deal is wrong, and Topolog is built around a different one: bring your own AI. Use our built-in authoring if you want one-click. But the entire planning language is public, the engine that validates plans is deterministic, and any model you already pay for, ChatGPT, Claude, Gemini, a local model, can author plans for you at no extra charge from us.
Why this is even possible#
The arrangement only works because of two architectural choices that most tools have not made.
First, plans are text. Every Topolog plan is a program in TOL, a small total language (the case for that is its own article). LLMs are spectacular at reading and writing structured text and hopeless at operating someone else's drag-and-drop canvas. A tool whose plans live behind a mouse has no honest way to let your model in; a tool whose plans are text has no honest way to keep it out.
Second, correctness is checked, not vibed. The TOL validator enforces roughly eighty structural invariants, deterministically, in milliseconds, on every edit: unreachable work, non-exhaustive branches, dependency cycles, container views that disagree with task-level edges, oversized decompositions. This is the piece that makes third-party authoring safe. The model drafts; the engine judges. The judge is not an AI, it does not hallucinate, and it treats our AI and your AI with exactly the same severity.
The forecast works the same way. Completion distributions, critical paths, and cash trajectories are computed by a seeded Monte Carlo engine from the plan's structure. No model guesses your odds. Which means the question "is the AI any good?" stops being about trust and becomes empirical: did the plan validate, and do you agree with the structure it wrote?
The workflow, concretely#
The whole loop takes about five minutes the first time:
- Get the handbook. The complete language reference is one markdown file, written to be pasted into a chat: topolog.co.uk/tol-handbook.md. Download it, or just give your model the link.
- Ask for a plan. A prompt that works well: "Read this handbook. Write a TOL plan for [your goal], with realistic estimates and cv values, milestones, an outcome-gated ending, and sentinels for success and failure."
- Paste it into the Source tab. The validator runs as you paste. A first attempt from a capable model typically lands with a handful of errors, each pointing at a line and naming the violated invariant.
- Feed the errors back. Paste the error list into the same chat. This is the step that feels like magic and is actually just a compiler loop: models are very good at fixing named, located errors in a language they have the spec for. One or two rounds usually reaches zero.
- Read the forecast. The moment the plan validates, the spectrum, schedule, and money views light up, computed by the engine, not asserted by the model.
The error messages are designed for this loop. A model that writes a dependency cycle gets back the named edges that close the loop; a branch that does not cover both values of an outcome gets the missing case spelled out. Errors as a conversation format turn any LLM into a competent TOL author, because the part LLMs are weak at (global structural consistency) is exactly the part the engine checks.
What it costs (and what it does not)#
The built-in Express and Structured builders run on pooled workspace credits: a plan build costs a fixed, published amount, the pool is shared by the team rather than priced per seat, and you spend it only when you actually build. The BYOAI path costs nothing on top of your existing subscription to whichever model you prefer. Validation, forecasting, scheduling, and every view in the product work identically on a plan regardless of who or what authored it.
| AI add-on (typical PM tool) | Bring your own AI | |
|---|---|---|
| Price | Per seat, per month, always on | Your existing LLM subscription; our credits only for built-in builds |
| Model choice | Whatever the vendor picked | Whatever you prefer, including local |
| Output | Suggestions inside a proprietary UI | A text artifact you can read, diff, and keep |
| Correctness | The same model marking its own homework | A deterministic validator with eighty-odd invariants |
| Lock-in | The work lives in their AI features | The work is a portable text file |
The honest trade-off: the BYOAI loop is a power-user workflow. It involves a chat window and pasting, and the first round trip teaches you what cv values and sentinels are. People who want none of that click the Express button and get a built, validated, forecast plan from one paragraph of goal description. The point is not that everyone should bring a model; it is that nobody should be locked into ours.
Why we are comfortable with this#
A reasonable question: if your model can author the plans, what exactly is Topolog selling? The answer is everything the model cannot do. A language with termination guarantees. A validator that knows what a well-formed plan is. A seeded Monte Carlo engine for the odds. A scheduler that knows your painted availability and your team's capacity. The execution loop that learns your real pace from completed work and recalibrates the forecast.
Drafting plans was the part AI commoditised. We would rather you spend that commodity wherever it is cheapest for you, and pay us for the part that has to be correct: the engine. The handbook is at /docs/tol-handbook, the deeper BYOAI guide at /docs/bring-your-own-ai, and your first prompt is one paste away.