Jungle Labs
ENPL Polski
Get in touch

AI agents do the legwork. You keep the final say.

We build agents for defined jobs: analysing contracts, answering customer emails, booking meetings and preparing quotes. Each one works from documents you approve, and anything that matters waits for a person before it goes out.

Customer service inboxesContract and document reviewQuotes and proposalsMeeting and appointment booking
01Before and after

What moves to the agent.

Today: Your team writes the same order and delivery answers from scratch, one email at a time.
With an agent: The agent drafts each reply from your order data and policies. A person approves or edits it.
Today: A manager reads every returned contract in full to find what the other side changed.
With an agent: The agent lists each change against your standard terms, with clause numbers, for the manager to decide.
Today: Quotes wait days while someone gathers prices, specs and lead times from different files.
With an agent: The agent drafts the quote from your price list and product sheets. Sales checks it and sends it.
Today: Booking a call takes several emails to find a time that suits everyone.
With an agent: The agent offers free slots from the right calendar, books the call and sends the invite.
Today: New staff ask colleagues where a procedure is written down, or guess.
With an agent: Staff ask the agent and get an answer from your documents, with the source linked.
02What we build

What the agents do, and how they're checked.

Each agent does one defined job, works from sources you approve and hands anything uncertain to a person.

  • Customer email assistant

    It reads each email, looks up the order or account and drafts a reply from your policies and past answers. A person approves every draft until you decide routine replies can go out on their own.

  • Contract review

    It compares a contract, or the other side's changes, with your standard terms and your list of accepted positions. It lists each difference with the clause and wording, and a person decides every point.

  • Quote drafting

    It matches an emailed request to your catalogue, applies your price list and discount rules, and drafts the quote in your template. Anything it cannot match is flagged for sales.

  • Meeting booking

    It answers requests for a call, offers times from the right person's calendar and sends the invite with a short agenda taken from the conversation.

  • Grounding in your documents

    The agent answers from documents and data you approve, such as policies, price lists and order records, and shows the source of each answer. When they don't cover a question, it is built to say so and pass it to a person.

  • Approval steps and logs

    Where a mistake would be costly, we add a human in the loop: a person approves the step before it runs. Every action, approval and edit is logged, so you can see what the agent did and who approved it.

03Workflows

Three agents, step by step.

The companies and volumes are invented, but the steps and the points where a person takes over are the ones we build.

  1. Example scenario

    Order and delivery questions in a customer inbox

    A customer emails to ask where an order is, or how to return an item.

    1. The agent finds the customer and the order, then pulls the status and tracking from your systems.
    2. It drafts a reply from your returns policy and style guide and cites the policy section it used.
    3. Complaints, refunds above your limit and low-confidence drafts go straight to a team lead.
    4. Routine drafts wait in a review queue, where a person approves or edits each one before it is sent.
    5. Every draft and edit is logged, and the edits become new test cases.

    Routine replies take a minute to check instead of five to write. The team gets back about 10 hours a week for the cases that need a person.

    Assumes 60 emails a day over 5 days, half of them routine. A reply takes 5 minutes to write by hand and 1 minute to review as a draft.

  2. Example scenario

    Returned contracts checked against your standard terms

    A customer returns your contract with tracked changes, or a supplier sends their own terms as a PDF.

    1. The agent compares each clause with your standard terms and your list of accepted positions.
    2. It lists every difference with the clause number, your wording and theirs.
    3. Each difference is marked as accepted before, new or outside your limits, with the rule it applied.
    4. A manager or your lawyer decides each point and writes the reply.

    The reviewer starts from a list of differences, each with the source text beside it, and gets back about 12 hours a month. Every decision stays with a person.

    Assumes 12 contracts a month, 90 minutes to read and compare each by hand, and 30 minutes to check the agent's list.

  3. Example scenario

    Quote requests turned into draft quotes

    A prospect emails a request for quotation listing products, quantities and a delivery date.

    1. The agent matches each item to your catalogue and flags anything it cannot match.
    2. It applies your price list, volume discounts and current lead times from the stock system.
    3. It drafts the quote in your template and lists open questions for the salesperson.
    4. The salesperson checks and sends it. Discounts beyond your rules need a manager's approval.
    5. If the prospect asks for a call, the agent offers times and books it in the salesperson's calendar.

    A quote now takes a 15-minute check, down from 45 minutes of assembly. Sales gets back about 8 hours a week for follow-up.

    Assumes 16 quote requests a week, 45 minutes to build a quote by hand and 15 minutes to check the agent's draft.

04Models and tools

Models chosen for each task.

We test candidate models on your own past cases and choose per task, weighing quality, cost and where the data is processed.

Models
OpenAIAnthropicMistralGoogleOpen-weight models
Orchestration
n8nMakeCustom code
Search and storage
PostgreSQLpgvectorSupabase
Where your team uses it
EmailSlackMicrosoft TeamsYour helpdeskA simple web panel

Names are for reference, and we choose per project. Model and tool accounts are opened in your company's name, so you see usage and costs directly and keep everything we build.

05Timeline

How an agent earns its place.

  1. 011–⁠2 weeks

    Audit

    We gather real past cases and the answers your team gave, agree what good looks like and mark the steps that need a person.

  2. 022–⁠4 weeks

    Pilot

    We build the agent and test it on those cases. Then it runs in draft mode, where every output goes to a person before anything happens.

  3. 032–⁠6 weeks

    Deploy

    We set approval thresholds with you, add logs and alerts and train the reviewers. Routine steps run alone only where the pilot results support it.

  4. 04Ongoing

    Operate

    We review samples of its work, retest when your documents or the models change and report what it handled and what it passed to people.

06Questions

Fair questions about agents.

What happens when the agent gets it wrong?

Agents can be wrong, so each one is built to hand over when it is unsure. Low-confidence answers, questions your documents don't cover and anything above the limits you set go to a person. Every action is logged, so a mistake can be traced and corrected.

Can it make up an answer?

It answers from the documents you approve and shows the passage behind each answer, so a reviewer can check it. That lowers the risk of an invented answer but does not remove it. Before launch we also test it with questions your documents can't answer, to confirm it says so.

Do the model providers train on our data?

No. We use business and API plans whose terms exclude training on your data by default. Where a provider leaves it to a setting, we switch training off. Providers still keep request data for a limited time to detect abuse, typically 30 days. Where a project needs it, we apply for zero data retention. Data that must be processed in the EU goes to EU-region services or open models hosted in the EU. How we handle data.

What if a model provider changes its prices or terms?

The agent's instructions, test cases and documents belong to you and are kept separate from any one model. Moving to another provider means rerunning the tests and adjusting the instructions. Under an Operate & improve arrangement, that work is included.

How is an agent priced?

The build is a fixed fee, agreed before we start. Model usage is billed to your own account, so you see the real running cost. Monitoring and improvement afterwards is a monthly arrangement, if you want it. See how we work.

What do you need from us?

We need access to the documents the agent should use and a set of real past cases with the answers your team gave. We also need one person who can judge a good result, for an hour or two a week during the pilot.

Will customers know they are dealing with an AI agent?

Yes. EU law requires that people are told when they're dealing with an AI system. An agent that writes to customers on its own says so in its first message and in every email. It also names the company it writes for. How we handle the AI Act.

Start with the inbox that never empties.

Book an audit. We read a sample of your real requests, estimate what an agent could handle and show you where a person should stay in charge.

Book an audit