# AI speed, with a person approving every stage

> AI-Accelerated is Tog Lab’s way of building software with AI: AI agents do the heavy lifting across seven stages, and a person approves each one before the next begins.

URL: https://www.toglab.co/en/ai-accelerated.html

[Português do Brasil](https://www.toglab.co/pt-br/ai-accelerated.html.md) · [Português de Portugal](https://www.toglab.co/pt-pt/ai-accelerated.html.md) · [Español](https://www.toglab.co/es/ai-accelerated.html.md) · [Italiano](https://www.toglab.co/it/ai-accelerated.html.md)

[Tell us about your project](https://www.toglab.co/en/contato.html) See how it works

Faster than the traditional cycle, with a record of who approved every decision.

## How it works, from request to production

The same cycle repeats at each of the seven stages. AI does the heavy lifting and a person decides whether the stage is ready.

Seven approval gates, one at the end of each stage. At each gate, a named person with a defined role reads the stage document and decides whether the work moves forward. AI does not ship anything on its own.

See what happens at each stage

## The seven stages

Always in the same order. Pick a stage to see what AI does, who approves it and the document that gets recorded.

At each stage, the badge shows who approves before the next one starts.

AI does

Approval gate

What gets recorded

Sample project

## Inside AI-Accelerated

The technical pieces that keep AI working within your company’s standards. Pick a piece to see what it does. The technical name appears next to it, for anyone who wants to go deeper.

## How far AI goes today

Four levels of autonomy. AI-Accelerated works up to the third, always with an approved plan and review by a person.

AI edits files outside the approved plan, or in sensitive areas such as payments, authentication and personal data, only with explicit approval from the Tog Lab tech lead. The autonomous agent is Tog Lab research and is not part of any commercial proposal today.

### AI-Accelerated inventory

A proposal includes only what has already been validated in client projects. The rest we use in our own projects or are still studying.

Inventory as of October 2026.

## The traditional cycle, shortened with AI

A case study measured by Tog Lab

We compared the same deliverables of a SaaS product under two models: the plan for the traditional cycle and the execution with AI-Accelerated. The domain was well defined. On legacy systems, the gain is usually smaller.

163 daysunder 5 days

of work across the three deliverables measured in days: feature plan, authentication and payments.

Fewer hours of work for the same deliverable mean lower project cost. On every project, we measure the starting point and the actual gain before promising any number.

Each row compares both cycles for the same deliverable. Bars are roughly to scale in hours of work.

## Governance

Every decision has an owner and a record, and AI-Accelerated does not depend on a single AI model.

References used to design the governance model: NIST AI RMF, OWASP Top 10 for LLM Applications and ISO/IEC 42001. They guide the design and do not indicate Tog Lab certification.

[Read: The driverless car and AI-assisted software development](https://www.toglab.co/en/insights-carro-sem-motorista.html)

## Where AI-Accelerated runs

The method stays the same. What changes is who leads and where it runs.

## See AI-accelerated development on one of your business workflows

Someone from business and someone from engineering will talk with you. We reply within 1 business day.

[Tell us about your project](https://www.toglab.co/en/contato.html) [hello@toglab.co](mailto:hello@toglab.co)

## More on this page

- In Tog Lab’s own projects — The seven stages with human approval — Harness with context, rules and logging — AI-accelerated screen prototyping — Team training on AI-Accelerated
- In research — Agent that acts as PO or PM — AI-assisted mobile development — Legacy system rewrites — Autonomous workflow, with no person in between

- Context — Discovery consolidated into a single source of business rules, risks and decisions. — Organizes interviews and documents and proposes the business rules. — The business owner on your side validates the rules and goals. — The versioned Domain Context in your repository. — version 1.1 — The document with the goals, business rules and risks. Every later stage starts from it.
- Requirements — Requirements, user stories, priorities and risks. — Drafts requirements, user stories and risks from the context. — The business owner decides what is in and what is out. — The PRD with priorities and the approval date. — version 1 — The list of what the system needs to do, in priority order, with user stories from the people who will use it.
- Architecture — Architecture and infrastructure with the choices explained. — Proposes architecture options with cost, pros and cons. — The Tog Lab tech lead chooses and records the reason. — The architecture document and every recorded decision (ADR). — A recorded technical decision: what was chosen, the options considered and the reason.
- Specification — Flows, errors and acceptance criteria for each feature. — Details flows, error messages and acceptance criteria. — The tech lead reviews and approves the specification. — The specification for each feature, with its acceptance criteria. — version 1 — The detail of one feature: how it works, what happens when there is an error and when it is done.
- Prototype — Screens generated from the specification and validated with the people who will use them. — Generates clickable screens from the specification. — Design and the business owner validate them with end users. — Every prototype version and what changed in it. — version 3 — The record of the prototype test with end users: what changed from one version to the next.
- Infrastructure — Setup and deploy described step by step. — Prepares the step-by-step guide and the deploy configuration. — The tech lead runs the deploy and confirms system health. — The step-by-step guide and the log of every deploy. — version 1 — The deploy steps, so every deploy is repeated the same way.

- AI model and key — AI-Accelerated calls whichever AI model your company chooses, through the provider and in the region that make sense, with the key and the invoice in your company’s name. Switching models is a configuration change. It does not require a new project. — Excerpt of the configuration that defines which AI model is used and in whose name. — provider: the one your company chooses — region: your company’s region — key: client account — id: the model your company chooses — model switch: only this block changes
- Harness — The structure around the agents: business context, company rules, the tools each agent may use, checks and logging. The name comes from a horse’s harness: the model provides the power, and the harness directs that power into useful work, within limits. It is what keeps AI working to the company’s standards instead of answering on its own terms. — The harness configuration: the context, the agents, the rules and who reviews.
- Business context — Domain Context, context engineering — Every project starts with a Domain Context: rules, risks and decisions in a single source. Each stage gets only the context it needs, and the documents from one stage (PRD, architecture, specifications) feed the next. — A business rule from the Domain Context, which every agent receives. — The data belongs to the grower. The — co-op sees only what they authorize. — Weak signal in the field. — approved by: product, client
- Specialized agents — agents and subagents — One agent per type of work: requirements, architecture, code, tests and documentation. Each has its own instructions, tools and permissions, and a reviewer agent checks the others’ work before the person does. — What each agent can use and where it can write.
- Guardrails — guardrails and permissions — Rules that block, before anything runs, what must not happen: destructive operations, unlicensed dependencies, data access outside the role and changes to sensitive areas without explicit approval. — The log of what the rules blocked before anything ran. — guardrail no-destructive-ops — blocked DROP TABLE in production — guardrail licenses — blocked unlicensed package — data schema only, no rows
- Automated checks — hooks, TDD, evals — Hooks run at every agent step: tests written before the code (TDD), evals that measure response quality, regression and security standards before any merge. — The result of the checks that run before every delivery. — hook pre-commit ok — tests 42 passed — evals no regression — security OWASP LLM ok — review awaiting the person
- Human approval gate — human in the loop — At the end of each stage, the designated person reads the artifact and approves, adjusts or rejects it. Every pull request is reviewed by a Tog Lab professional before it is merged. Nothing goes from an agent straight to production. — Who approved each stage, and what is still waiting. — stage 01 Domain Context v1.1 — stage 02 PRD v1 — stage 03 HLD v1 — awaiting tech lead
- Audit trail and repository — ADR, changelog, observability — Architecture decisions become ADRs, every change goes into the changelog and every document is versioned. Token usage per stage is visible. Everything lives in your company’s repository from the first commit. — A recorded decision, with the options, the reason and who decided. — ADR-07 Website CMS — options Payload, Storyblok — decision Payload — reason data in our database — by tech lead — tokens for this stage: in the project dashboard

- Level 1 — Autocomplete — AI suggests the next line and the engineer accepts or discards it. — Everyday use
- Level 2 — Edit a snippet — AI rewrites a function on request, and the person reads the diff before accepting. — Everyday use
- Level 3 — Edit files from a plan — AI changes files based on an approved plan, with guardrails, tests and review by a person. — Current limit
- Level 4 — Autonomous agent — Would plan, execute and ship with no person in between. Requires evals and permissions still being validated. — In research

- You bring the request — A feature, a new product or an improvement. We start from your business context.
- The agents prepare the stage — AI agents write the document, the prototype or the code for that stage, within your company’s rules.
- A person reviews and approves — Nothing moves on without approval. If something is not right, the person adjusts it or sends it back for the agents to redo.
- The next stage begins — There are seven stages up to production. Every approval is recorded in your repository.

- Who approves on your side — Your company’s business owner. — Business goals and rules — Priorities for what comes first — Prototype tested with end users
- Who approves at Tog Lab — The tech lead and the Tog Lab team on the project. — Architecture and technical decisions — Specification of each feature — Code, tests and production deploy