Skip to content
AI AGENTS

Ready-made or custom AI agent? When an off-the-shelf agent is enough, and when you need approval steps

A ready-made AI agent is enough when the task is standard, a person checks the result, and nothing is paid, released or sent on its own. You need a custom agent when it must work inside your own systems. It must also stop for a named person before money moves or code is released.Akaari builds one from US$9,550. If a ready-made agent does the job, use it.

Prices exclude VAT.

When a ready-made agent is enough

A ready-made agent fits when all four of these are true.

  • The task is the same across companies. Screening CVs. Scoring leads. Matching invoices to purchase orders. Triaging help desk tickets. Drafting proposals.
  • A person reads the output before anyone acts on it. The agent suggests. Your team decides.
  • The agent writes little or nothing into your own systems. It works inside its own screen. Someone copies the result across.
  • Usage pricing fits your volume. You pay for what you run. The monthly bill makes sense at your real numbers.

Ready-made agents are usually sold as subscriptions, often with a sandbox trial first. Some agent marketplaces offer ready-made agents for jobs such as purchasing, finance, sales and HR. You often start on trial credits and move to a paid plan after that.

If that is your case, start there. You may not need us.

When you need a custom agent

You need a custom agent when one or more of these is true.

  1. The agent must act, and suggesting is not enough. It has to write into your CRM, ERP, billing or product.
  2. An action has consequences. Moving money. Deploying to production. Sending to a real customer inbox.
  3. The workflow is yours. Your rules, your exceptions, your approval chain.
  4. You need a record of who approved what. For your CEO, your auditor, or a regulated client who asks.
  5. You need a say in where models and data run. We assess this in discovery. We never assume it.

A custom agent costs more and takes longer than a subscription. It is worth it only when the workflow moves a number you care about. If AI cannot move the number, we say so before you pay to build.

For what the build covers, see an AI agent for one workflow.

What an approval gate is

An approval gate is a step where the agent prepares an action and then stops. A named person approves or rejects it. Named means a specific person or role, agreed before the agent goes live. "Someone on the team" does not count. Every approval or rejection the agent asks for is logged with who decided and when.

At Akaari, three actions always stop for a named person:

  • Money movement.
  • Production deploys.
  • Sending to a real inbox.

The agent can ask for a release but can't release on its own. Production access stays with a person.

An example gate list (example, not client work):

WorkflowWhat the agent doesWhere it stopsWho approves
Supplier paymentsMatches each invoice to its purchase order and drafts the paymentBefore the payment is releasedThe finance lead named for this workflow
Customer repliesDrafts the reply from the ticket and the account historyBefore it is sent to the customerThe support lead named for this workflow
Product releasesPrepares the release notes and asks for a releaseBefore anything reaches productionThe engineering lead named for this workflow
Run card · codeWaiting on you

Build. Test. Fix. Deploy after a person approves.

Demo: finish the build and tests. Deploy waits for a person.

Production access stays with a person. Approve deploy.

Tokens
13350 / 15000
Minutes
42
Retries
1
  • TaskBrief: Task brief. Goal, objects, metric, and the questions still open.
  • TestPlan: Analyze note. Web only. This hop does not deploy.
  • ShipPackage: Patch. The Start page and the leads client. Staging shape.
  • TestPlan: Test plan. Cases are written. Runners: desktop, browsers, mobile, and API.
  • TestPlan: Desktop UI. Submit and validation.
  • TestPlan: Browser matrix. Passed.
MANAGER AI · LOW CONTROLBuild. Test. Fix. Deploy after…Digest2 waiting on youExceptionsPulse quiet unless money movesWatch
Akaari's internal system, not client work.

The run card shows how a gate works inside Akaari's own system. It does not prove results for any client. See how the approval gates work and how Akaari agents work.

Side by side: ready-made vs custom

This table compares the two types. It does not describe any one provider.

Ready-made agentCustom agent with Akaari
What it doesA standard task that looks the same across companiesYour workflow, with your rules and exceptions
Where it actsMostly inside its own interfaceInside your systems: CRM, ERP, billing or product
ApprovalsWhatever the product offersA named person for each action that moves money, deploys or sends, with a log of who decided and when
Pricing modelUsage packagesA build fee, then a monthly running cost plus model and hosting at cost
Time to startOften a trial you can start soon4 to 8 weeks of build after paid discovery
After go-liveThe provider60 days of free fixes after launch. Then a monthly fee for monitoring, fixes, small changes and training, which you can stop at any time. The agent and code stay with you
Who owns itThe provider, under its own termsYou. Code, prompts, test cases and accounts

Try ready-made first: what to test in the sandbox

  • Accuracy on your own cases. Run it on your real cases, including the awkward ones.
  • Exceptions. When it hits a case it does not understand, does it stop, guess, or hand it to a person?
  • Writes. Can it write into your systems? Which ones, and with what access?
  • Approvals. Who can approve an action? Can that be one named person for each kind of action?
  • Cost at your volume. Price it at your real monthly volume, after the trial credits run out.
  • Data. Where it is stored, who can see it, and how long it is kept. Ask the provider to put this in writing.

If it passes, keep it. If it fails on a gate or an integration, that is the brief for a custom build. Write down where it failed. That list is the most useful thing you can bring to discovery.

What a custom agent costs and how long it takes with Akaari

  • One workflow agent: US$9,550 to US$21,800 to build, 4 to 8 weeks after paid discovery. Prices exclude VAT.
  • Fixes after launch: free for 60 days.
  • Running cost: US$700 to US$2,050 a month. It covers monitoring, fixes, small changes and training for your people. AI model and hosting usage is extra, at cost in your account. Stop it at any time. The agent and code stay with you. Prices exclude VAT.

What moves the price:

  • The systems the agent must read from and write to, such as your customer records (CRM), billing or payment provider.
  • The actions with side effects, and the gates each one needs.
  • Volume, and the model that fits it.
  • The accuracy bar the workflow needs.
  • Where data must stay.
  • How much monitoring and change it needs after go-live.

Every build starts with paid discovery. It takes 2 weeks and costs US$3,250 to US$5,450. The fee is credited in full if the build is signed within 30 days of the day the brief is delivered. The brief is yours to keep, even if we don't build. The brief sets the exact price.

Using both

Most companies won't pick one side. Keep ready-made agents for the standard tasks. Build one custom agent for the workflow with consequences.

If the workflow is really an internal app with no AI in it, read replacing Excel and WhatsApp with an internal app. If an agent or automation someone else started has stalled, start with a takeover audit.

See also AI strategy, what an MVP costs and how we handle access and data.

Questions people ask

When is a ready-made AI agent enough?

When the task is standard across companies and a person reviews the output before anyone acts on it. It also has to work without writing into your own systems or moving money. Try it in a sandbox on your own real cases first.

When does a company need a custom AI agent?

When the agent must act inside your systems, the workflow follows your own rules, or an action has consequences. Paying, deploying and messaging a real customer are actions with consequences. Each one must stop for a named person, and the decision is logged.

What is an approval gate in an AI agent?

A step where the agent prepares an action and stops until a named person approves it. At Akaari, money movement, production deploys and sending to a real inbox always stop for a person.

Can we start with a ready-made agent and build a custom one later?

Yes, and often you should. The trial teaches you things such as where it fails and which approvals you need. That becomes the input for discovery.

How much does a custom AI agent cost with Akaari?

US$9,550 to US$21,800 to build one workflow agent. Plus a running cost of US$700 to US$2,050 a month, and AI model and hosting usage at cost. Prices exclude VAT.

How long does a custom AI agent take to build?

4 to 8 weeks after paid discovery. Discovery itself takes 2 weeks.

Who owns the custom agent?

You do. Everything built for you belongs to you: the code, the prompts, the test cases we check it against and the accounts. Model and hosting usage runs in your own account, at cost.

What happens after the agent goes live?

Fixes are free for 60 days after launch. Then the monthly fee of US$700 to US$2,050 covers monitoring, fixes, small changes and training for your people. AI model and hosting usage is at cost. Prices exclude VAT. You can stop it at any time. The agent and code stay with you.

NEXT

Not sure which side your workflow falls on?

Tell us the workflow and the action you would want a person to approve. A person from Akaari, not a bot, makes first contact within 30 minutes of your request. The first step is a fit conversation. It is free. It is not a quote and not a contract. Paid discovery starts only when both sides agree.