You can also browse, search and filter all release notes directly in the application → Open release notes in TAI.

​

Major New Feature.
​

Private preview has ended — TAI is now generally available.

Thank you for building with us through the preview — this is just the beginning.

​

Major New Feature.
​

A ready-to-use built-in skill. The Travel Regulations Legal Advisor works out which regulation governs a situation, and how they stack — more than one usually applies.
​

Five bodies of law ship with it, each as the verbatim official text plus a Nezasa-authored operator brief that says what it means in practice:

The agent reasons from the briefs, quotes the official text, attributes every source, and always closes with a not-legal-advice notice. Nothing to set up — the reference material comes with the skill.
​

Add your own documents next to it. Create a custom agent, attach the skill, then add your company's own material as Knowledge — booking terms and conditions, cancellation and refund policy, baggage rules. The agent then answers against both: what you promised the traveller, and the regulation that governs it. That combination is where the useful answers are — your terms rarely tell the whole story, and the regulation alone doesn't know what you sold.
​

Going further: add Fetch TripBuilder Itinerary by Ref-id (paired with Cockpit API — Customer Care). Mention a booking reference and the agent pulls the live itinerary and answers against it — cancellations, delays, refunds, data requests.
​

The Travel Regulations Legal Advisor skill, with its required knowledge collection attached.

​

Major New Feature.
​

Agents can now search knowledge you attach to them — files, snippets, or a crawled website — and cite what they find, grounding answers in your own material instead of general knowledge.

Attach knowledge or a collection to any agent from its Definition section, or to a skill so every agent that uses the skill brings the same grounding along automatically.
​

The Knowledge catalogue, with a Partial index badge visible in context.
​

A Stale badge, with the notice explaining how long it's been since the last successful index.
​

A Partial index source's detail page, with the skipped-pages breakdown.
​

Preview search, showing the ranked chunks an agent would retrieve with their match type and relevance.

​

New Feature.
​

The TAI Assistant is now wired to the built-in TAI Help Center, so it can answer questions about TAI itself — what agents, skills, connectors, and knowledge are, and how to use the Workbench — citing its sources along the way.
​

This applies where you work in TAI — Chats and the Terminal. Embed the agent in your own site through the Chat UI SDK, or call it from your own backend, and it drops the help center, the prompt coaching and every mention of TAI in its greeting: your visitors get a general assistant, not a tour of our product.

​

New Feature.
​

A new built-in Prompt Coaching skill, attached to the TAI Assistant, helps you write, improve, or debug the system prompt for an agent you're building: it asks what the agent is for, walks a clarity checklist, and hands back concrete before/after rewrites for you to paste in and test yourself.
​

It is the conversational way in — think a prompt through, out loud, before you commit to it. When you are ready to act on it, Improve on the agent's own page is the stronger tool and available to everyone who can author an agent: it rewrites the stored prompt rather than the excerpt you pasted into chat. Company administrators can go further with an Advisor run, which evaluates the agent's real conversations and proposes changes you can apply.

​

New Feature. Only Company Admins.
​

Corporate design now has a shortcut: point it at your website, or upload a logo file, and TAI proposes matching colours, fonts, and a corner radius pulled from what's already there. Nothing is saved until you review the fields and click Save changes, and anything TAI can't confidently derive is simply left for you to fill in by hand. Deriving from a website respects your company's Web access settings — if a site is blocked or web access is off, Settings → AI & Data → Web access is where to fix that.
​

Point Derive corporate design at your website — nothing saves until you review and click Save changes.

​

New Feature.
​

Ask a TripBuilder agent which distribution channels your tour operator has and it will name them. Previously it could work within your channels but could not enumerate them, and would tell you the list was outside what it could see.
​

The list it gives you is the authoritative one, so it no longer has to piece an answer together from whichever channels happened to appear on the bookings it looked at.
​

The agent has also stopped overstating your own access. Until now, anyone signing in through the Workbench was told they had access to all channels of their tour operator — true for most people, but wrong for anyone whose account is restricted to specific channels, and it could lead the agent to read a legitimately narrower result as missing data. It now states your channel scope only when your sign-in actually proves it.
​

No action needed. The channel list appears once the supporting TripBuilder platform release is out; until then agents behave as they do today.

​

New Feature. Only Company Admins.
​

Acting on an advisor run used to mean reading technical findings and rewriting the prompt yourself — and some findings pointed at a prompt proposal that was not there. A finished conversation evaluation or report run analysis now carries a Propose an improved prompt from this analysis action. One click drafts a concrete replacement system prompt — plus any skill changes the findings justify — with the same diff preview and Apply you already get on a proposal, so you can see exactly what would change before anything is written.
​

You can steer it with a sentence ("focus on the data-source findings"), and it always works from the agent's definition as it stands right now, not as it stood when the evaluation ran. Older runs whose proposals went missing can be recovered this way: the findings survived, so a proposal can be drafted from them.
​

Two things to know: it is not offered on built-in agents, whose prompt is managed by Nezasa — copy one into your company first and run the advisor on the copy. And each draft is its own advisor run, so it is billed and counts toward your daily limit.

​

Major New Feature. Only Company Admins.
​

Company admins can now set their own brand for every report and export TAI generates — a new Corporate design page under Settings → Workspace covers seven brand colours (accent, ink and surface, each with a softer variant, plus a border colour), a five-colour severity palette for risk-style reports, font and layout choices, and the tone of voice reports should read in. Leave any field empty to keep TAI's shipped default — a colour goes back to its default via the clear button beside its picker — or use Reset to defaults to clear every override at once. A report TAI produced as a formatted document keeps the look it was created with; one stored as plain text is styled at download or send time, so a change shows up on those earlier reports too.
​

The new Corporate design settings page, with brand colours and the severity palette — each field showing its default and what it paints.

​

Improvement.
​

The built-in agent nezasa/generic-assistant is now nezasa/tai-assistant. The old slug still resolves for any existing Chat UI SDK embed pinning it, and historical chat sessions, usage events, and reports are unaffected — only the name and slug changed.

​

Improvement. Only Company Admins.
​

The Advisor is now the Prompt Advisor. It has also moved out of AI Platform to sit with Settings and Documentation — help belongs with help. What it does is unchanged; only its name and its place in the menu.
​

Everything it covers, in one place:

Runs use your company's AI usage allowance and are capped per day.
​

A finished run: each proposal says why it was suggested and applies on its own, and the whole analysis can be turned into a rewritten prompt. Prompt Advisor now sits below Settings and Documentation.

​

Major Improvement.
​

AI help for writing an agent's system prompt is built into the editor, and now goes by one name — the Prompt Advisor. Under the System Prompt field, on any agent you create:

Every one of them proposes; nothing changes until you apply it and save.
​

The five actions sit under the System Prompt field. Here a rewrite has just been applied — still unsaved, and one click from Undo.

​

Improvement.
​

The Billing Center guide now covers all three credit families — AI compute, connector calls and storage — what makes each go up, and the rate for each, including the model categories (Fast, Balanced, Most capable) and the three connector tiers.
​

It also separates what you consumed from what you are invoiced for: Included credits never reach an invoice, and a Billable figure is what your usage costs at the standard rates rather than an amount owed.
​

A chat's and a report's credit figure now also says what it is: a consumed total, billable and included credits together. The split between the two stays in the Billing Center.

​

Improvement. Only Company Admins.
​

A finished report gains an Analyze this run action in its header. It hands the run to the Report Run Advisor — which reads the run's outcome, its reasoning stream and the delivered report, then reports back on whether the agent met its contract, where it wasted effort, and what to change — and opens the advisor run so you can watch it work.
​

The advisor could already do this; there was simply nowhere to ask from. The button existed only on the run detail page, and a report that succeeds never stays there — the library links it to its report, the run page redirects to it, and the launcher opens it the moment the run finishes. The action now lives on the page you actually land on.
​

Analyze this run, in the header of a finished report.
​

The advisor run it opens, reporting on how the agent handled the run.

​

Fix.
​

A report that corrected its figures mid-run attached both the draft and the correction, with no hint which was final. Re-rendering an export now replaces the earlier file.

​

Fix. Only Company Admins.
​

A model generation named with a whole number (Opus 5, Sonnet 5) sorted below its own 4.x predecessors, so Balanced and Most capable kept serving the older model.

​

Fix.
​

A connector call against a slow or unresponsive upstream system could leave the rest of that request unable to save anything, so what reached the agent, and the person waiting on it, was a generic server error. The call's own record was usually lost with it, hiding exactly the calls most worth looking into. Two things changed. Each call now keeps to its configured time limit as a whole, where before an upstream that answered a trickle at a time could run far past it. And when a call does have to be cut off, the failure stays contained to that call: the agent gets an error it can act on and move past, and the call is recorded in the connector call log like any other.

​

Fix.
​

Every weekly bucket runs Monday to Sunday, but the axis numbered it the US way — counting Sunday-start weeks from January 1. The two disagree around the new year: the bar for the week of Mon 4 Jan 2027 read W2 when it is ISO week 1, and the labels stayed one off for most of the year. Analytics, Billing and Usage Stats now all print the ISO week number, so a weekly figure lines up with any other ISO-week report. Labels only — the bars themselves never moved.

​

Fix. Only Company Admins.
​

Nezasa's built-in agent prompts are proprietary and hidden everywhere else in the product — advisor findings now match. On a built-in agent you still get the advisor's summary, issues, suggestions and reasoning, but not excerpts of that built-in agent's prompt. Two kinds of suggestion are withheld there too: a rewritten prompt, and a drafted skill — both restate the prompt itself. It also drops quotes taken from your own conversation with the agent, since the two share one field today; separating them is coming. Advisors on your own agents are unchanged.

​

Fix.
​

An HTML report — and the PDF exported from it — could contain plausible-looking but invented detail: promo codes, dates, destinations or per-row figures that were never in your data, alongside headline totals that were correct. The step that lays out the report was not being shown the results your agent had actually retrieved, so it filled the gaps itself. It now receives that data, and is instructed to render only what it was given and to flag anything it could not include rather than substitute a stand-in. Report footers also carry the real generation time instead of an invented one. Excel exports were never affected.

​

Fix. Only Company Admins.
​

An advisor proposal to attach or detach a skill used to identify it by its internal id — Skill: 9b718bc0-… — which told you nothing about what you were being asked to approve. It now shows the skill's name, falling back to the id only when the skill can't be resolved.

​

Fix.
​

Ask an agent what commission a TripBuilder booking earned and it now tells you what kind of figure it is, instead of presenting every figure as the amount paid.

​

Fix. Only Company Admins.
​