Back-Office Automation: Deel's Proven, Powerful Akai Launch

Back-office automation has a new and unusual entrant: a payroll company. Deel, the HR and payroll platform valued at $17.3 billion, has opened Akai to other businesses. Akai is the agent platform Deel built to run its own finance, compliance and HR operations. Deel first described it in May 2026 and offered early access, then relaunched it for general sale on 21 September with a money-back guarantee attached.

The back-office automation pitch is aimed squarely at operations teams rather than IT departments. An employee shows Akai a task once, or describes it in plain language, and the platform builds a workflow that a team of agents then runs, with human approval on anything that moves money or submits a filing. Deel says there is no need for a developer, an integration project or even a public API on the systems involved.

This article explains what Akai is, how its approach to back-office automation works, which workflows it targets, how it compares with robotic process automation and integration tools, what it costs and what it leaves unanswered. It closes with a checklist for evaluating any agent platform before you sign.

Table of contents

What Deel Launched: An AI Platform for Back-Office Automation
How Akai Approaches Back-Office Automation
Rules, Judgment and Approval in Akai’s Back-Office Automation
The Back-Office Automation Workflows Akai Targets
Akai vs RPA, Zapier and Browser Agents for Back-Office Automation
Back-Office Automation Security, Models and the EU AI Act
Back-Office Automation Pricing, Access and the Guarantee
What the Akai Back-Office Automation Launch Leaves Unanswered
How to Evaluate a Back-Office Automation Platform
References and Further Reading

What Deel Launched: An AI Platform for Back-Office Automation

Akai by Deel is described on its website, akai.run, as “the agent platform built for your operations” and as “multiplayer AI for enterprise operations”. The multiplayer label refers to its main design goal. A workflow built by one employee can be run, forked and extended by the rest of the team under their own logins, instead of every person building a private automation from scratch.

Two launches, four months apart

Deel announced Akai in a blog post dated 7 May 2026 under the headline “Manual work stops here.” At that stage it was an early access programme for “a select group of ops teams”. On 21 September, chief executive Alex Bouaziz posted “EXCITED TO LAUNCH: Akai” on X and opened it to any company with “hundreds of employees”. The second launch came with much larger performance claims from Deel’s own operations, which we examine in our companion piece on how Deel deployed 8,000 Akai AI agents internally.

Akai at a glance

Item	Detail
Maker	Deel, the HR, payroll and employer-of-record company
First announced	7 May 2026, as early access
General launch	21 September 2026
Who builds workflows	Operations staff, helped at first by a forward-deployed engineer
Reasoning models	Claude, ChatGPT or Gemini, chosen per workflow
Certifications	ISO 42001, ISO 27001, ISO 27701; GDPR compliant
Deel account needed	No; separate contract, infrastructure and pricing
Published price	None; the linked pricing page returned an error on 28 September
Guarantee	Full refund if Deel cannot automate 1,000 hours of work in 30 days

How Akai Approaches Back-Office Automation

Most back-office automation stalls at the same point: the systems operations teams use every day, such as bank portals, government filing sites and supplier platforms, have no usable API, and building one is an IT project nobody has time for. Akai’s answer is to learn the workflow from a person doing it and to work out the connections itself.

Step 1: show it the workflow once

The user records their screen while doing the task, narrating as they go, or describes it in writing or speech. Akai captures three things at once: the clicks on screen, the spoken context and the network traffic running behind the page. From that it maps the sequence of steps, the decision points and every system the task touches.

Bouaziz’s example was a payment reconciliation. An employee matches one messy wire transfer by pulling unformatted details from a bank portal, pasting them into a spreadsheet, checking invoices and payment history in NetSuite and raising a ticket in Zendesk. Akai “reads between the lines”, he wrote, and builds the workflow, including edge cases such as malformed invoice references, “without you writing a single regex”.

Step 2: review, approve and connect

Akai then generates the workflow as a script for the team to review before anything runs. The reviewer can change steps, decide which ones are locked and set how much control to keep. Akai also applies what it calls a built-in risk assessment framework, flagging workflows that need extra governance and adding approval steps by risk level.

Step 3: run, correct and improve

Once approved, the workflow runs on a schedule or trigger, at whatever volume the work requires, with as many agents as needed sharing a common context. Users correct it in plain English. Akai says every run adds to a store of the team’s rules, past cases and corrections, so each workflow’s back-office automation gets more accurate over time.

API first, browser last

The technical choice that separates Akai from screen-replay tools is the order in which it connects. Akai’s FAQ says it looks first for a public API. If none exists, it builds a custom connector from the network calls captured during the recording, effectively using the same unofficial API the website itself uses. It falls back to driving the browser only when there is no API at all. Those calls run under the user’s own credentials, so an agent has exactly the access its owner already has. If a call changes, the run stops and logs the failure rather than guessing, and Akai rebuilds the path.

Rules, Judgment and Approval in Akai's Back-Office Automation

For finance and compliance teams, the most important question about any back-office automation agent is not whether it is clever but whether it can be trusted with numbers. Akai’s design separates the two jobs explicitly.

Deterministic steps for anything that must be exact

Steps that have to be exact, such as a total, a tax rate, an account number or a payment reference, run on deterministic rules rather than a model’s judgment. The reviewer decides which steps are locked when approving the workflow. Deel’s argument is that locked steps carry no risk of a hallucinated number, because no model is deciding them.

AI reasoning for the judgment calls

AI reasoning is reserved for genuine judgment calls. Deel’s example is deciding whether a mismatched payment reference belongs to the invoice it appears to match. Anything ambiguous or high-stakes stops and goes to a person by Slack or email.

Approval gates on every write action

Write actions are gated by default. Anything that changes data, moves money or submits a filing waits for a named person to approve it. Every run is logged and can be replayed step by step. That is the control most back-office automation buyers will care about first, and the one to test hardest in a pilot.

The Back-Office Automation Workflows Akai Targets

Akai’s website lists 15 ready-made back-office automation use cases, most of them workflows Deel runs internally. They show where the company thinks back-office automation pays off fastest: high-volume, rules-heavy tasks spread across systems that do not talk to each other.

Finance and payments

The flagship is accounts payable reconciliation. The agent pulls invoices from Dynamics 365 or SAP, matches them against bank data from Citi, J.P. Morgan, HSBC and Wise, and prepares correcting entries for a person to approve. A month-end close agent pulls from NetSuite, reconciles against banks and builds the working papers. An expense review agent picks up the claims a first-line tool cannot clear, checking each against the country rules held in the HR system.

Compliance and financial crime

A sanctions triage agent screens customers and suppliers against ComplyAdvantage and OpenSanctions, sweeps crypto wallets through Elliptic daily, clears documented false positives and escalates real matches with the evidence attached. Deel says analyst time on documented false positives fell from five minutes to near zero, and an analyst still makes the final call on every true hit. An entity verification agent extracts company owners and directors from documents and cross-checks them against Dun & Bradstreet, OpenCorporates and Sumsub.

HR and IT administration

HR use cases include new-hire provisioning across Microsoft 365, Google Workspace and Slack, offboarding that revokes access across Entra ID, Okta and BambooHR, and an employment letter agent that checks Workday, drafts the letter and routes it through HelloSign for signature. A vendor risk agent keeps a register in Google Sheets in step with the active vendor list in Okta.

Logistics and supply chain

Deel also pitches its back-office automation beyond its own sector, with agents that prepare and submit customs documentation, monitor carrier portals for bookings and aggregate shipment tracking across a dozen portals. These pages carry no Deel performance figures, which is unsurprising for a payroll company, and logistics buyers should treat these back-office automation pages as templates rather than proof.

Use case	Systems named	Result Deel reports internally
AP and payment reconciliation	Dynamics 365, SAP, Citi, J.P. Morgan, HSBC, Wise	93.5% of cases end to end; 20+ days to minutes
Month-end close pack	NetSuite, HSBC, Wise, Airwallex	300+ entities, 25 reconciliation types a month
Expense review	HRIS country rules	Auto-clear rate 76% to 99%
Entity verification (KYB)	Dun & Bradstreet, OpenCorporates, Sumsub	About 10 minutes per case
Sanctions alert triage	ComplyAdvantage, OpenSanctions, Elliptic	False-positive handling from 5 minutes to near zero
Vendor risk register	Okta, Google Sheets	60 minutes of daily checks; review under 2 minutes
New-hire onboarding	Microsoft 365, Google Workspace, Slack	About 10 minutes per hire
Offboarding	Entra ID, Okta, BambooHR	1,300 to 1,500 tickets; 230 to 300 hours saved a month
Customs, carrier and shipment agents	Carrier and customs portals	No figures published

The chart below shows the share of cases Deel says Akai completes without a person, using only percentages stated on Deel’s pages or in Bouaziz’s post. The proof-of-location figure is 382 of 562 cases, from Deel’s FinCrime page.

Share of cases handled without a person, as Deel reports it (%)

Expense review, September page: 99

Payment reconciliation, use-case page: 93.5

Expense reimbursements, May launch: 90

Payment reconciliation, Bouaziz post: 85

Proof-of-location cases, first nine days: 68

Two of those bars describe the same payment reconciliation workflow at different figures, and the two expense bars come from different releases. The spread is a useful reminder that end-to-end rates depend on when and how they are measured.

Akai vs RPA, Zapier and Browser Agents for Back-Office Automation

Deel positions Akai against three familiar categories of back-office automation. Its FAQ argues that integration tools such as Zapier need pre-built APIs that most operations portals lack, that legacy robotic process automation such as UiPath is IT-led and “breaks when a portal changes”, and that browser automation tools replay the screen, break the same way and “can’t give you a clean audit trail”.

Approach	How it connects	Who usually builds it	Main weakness
Integration platforms	Published APIs and pre-built connectors	Business users or IT	Cannot reach portals without an API
Classic RPA	Scripted screen interactions	IT or a centre of excellence	Fragile when screens change
Browser agents	A model reading and clicking the page	Individual users	Variable results, weak audit trail
Akai	Public API, then captured network calls, then browser	Operations staff	Unproven outside Deel; no public price

Why Deel says legacy RPA breaks

The criticism of classic robotic process automation is familiar and partly fair. Bots that click through a user interface fail when a button moves or a page is redesigned, and they usually need specialists to maintain them. Akai’s claim is that calling a system’s underlying requests, rather than replaying the screen, means a portal redesign “usually doesn’t require you to do anything at all”. Our comparison of workflow automation, RPA and AI agents covers the general trade-offs in more depth.

The limits of Deel’s comparison

The comparison is Deel’s own, and it omits some trade-offs. Modern RPA suites also call APIs and include AI features, and integration platforms now offer agents too. Unofficial APIs captured from network traffic can change without notice and may sit awkwardly with a supplier’s terms of service. Akai’s promise to stop and rebuild when a call changes is sensible, but it still means a failed run to investigate. Buyers should ask how often that happens in practice.

Back-Office Automation Security, Models and the EU AI Act

A back-office automation platform that holds bank and government portal credentials is a security product as much as a productivity one. Cybersecurity teams should review it the way they would review any privileged access tool.

Which AI models run inside Akai

Akai does not have its own model. Its FAQ says “Claude, ChatGPT, and Gemini are the reasoning engines inside Akai”, that customers choose which one runs each workflow, and that Akai switches automatically if a provider goes down. Routine steps use lighter, cheaper models and complex decisions use more capable ones. Customers can also set the model and effort level for each workflow, and Akai shows the cost of every run.

Certifications and data handling

Deel says Akai uses an encrypted vault scoped to each user, role-based access control, a full audit trail and real-time alerting. It holds zero-data-retention agreements with all three model providers, so none of them trains on or retains customer data. Akai is certified under ISO 42001 for AI management systems, ISO 27001 for information security and ISO 27701 for privacy. Screen recording happens only during the session a user starts; after that, agents run through system connections with no ongoing recording.

Who carries the compliance risk

Every Akai use-case page carries the same disclaimer: using Akai “as a high-risk AI system under the EU AI Act and other applicable laws requires human-in-the-loop”, and the client, as deployer, is responsible for configuring that oversight. Annex III of the EU AI Act lists AI used to make decisions on employment, including promotion, termination and task allocation, among high-risk uses. Article 26 requires deployers to assign human oversight to people “who have the necessary competence, training and authority”.

An offboarding agent that revokes access after HR has decided a termination is processing a decision rather than making it, but the boundary is not always that clean. HR teams should check each workflow’s classification with counsel before deploying it in the EU.

Back-Office Automation Pricing, Access and the Guarantee

Akai’s back-office automation is sold separately from Deel’s payroll products, with its own contract. Deel says most teams have a first agent running within hours, and a dedicated forward-deployed engineer builds the first agents with the customer or for it.

No public price yet

Akai has not published a price. Its FAQ points to a pricing page, but that address returned a “page not found” error when we checked on 28 September. The offer is also pitched at executives “at a company with hundreds of employees”, which suggests enterprise-style contracts rather than self-service plans. For budgeting, our guide to business process automation cost sets out typical project pricing for comparison.

What 1,000 hours in 30 days means

Bouaziz announced an “Automation Guarantee”: “If our engineers can’t automate a thousand of hours of work in your first 30 days, you get a full refund.” The chart below shows 1,000 hours as a share of a company’s total monthly paid hours. It assumes 173.3 paid hours per employee per month, which is 40 hours a week multiplied by 52 and divided by 12.

1,000 hours as a share of total monthly paid hours, by company size (%)

100 employees (17,333 hours): 5.8

250 employees (43,333 hours): 2.3

500 employees (86,667 hours): 1.2

1,000 employees (173,333 hours): 0.6

For a company with a few hundred staff, the guarantee covers roughly 1% to 2% of all paid hours. That is a realistic target for one back-office automation programme’s first month, and it is the reason the guarantee is credible. The fine print, including how hours are measured and who verifies them, has not been published.

What the Akai Back-Office Automation Launch Leaves Unanswered

Akai arrives with more evidence than most back-office automation launches, because Deel has run it on its own operations. The evidence is still Deel’s own.

One named outside customer

The only external user Deel has named is Ostberg Sinclair & Co, whose operations manager, Frankie Limmer, said in May that building a first agent took under an hour and saved “over three hours a week”. That is a small-firm example. The large figures, such as more than 250,000 cases a month and more than 10,000 live agents, all describe Deel itself, and none has been independently audited.

Vendor concentration

Akai adds a vendor layer on top of three model providers. Workflows, captured connectors and the accumulated store of rules and corrections live inside Akai. Buyers should ask what can be exported if they leave, and how a change in a model provider’s terms would reach their workflows.

A crowded back-office automation market

Deel is entering a back-office automation market where Gartner warned in June 2025 that more than 40% of agentic AI projects would be cancelled by the end of 2027, and that many vendors were “agent washing” existing products. Akai’s production history at Deel sets it apart from pure start-ups, but it does not exempt it from that risk.

How to Evaluate a Back-Office Automation Platform

Whether or not Akai makes your shortlist, its design gives back-office automation buyers a useful set of questions. Ask every vendor the same ones and compare the answers side by side.

Question	Why it matters	Akai’s stated answer
How does it connect to each system?	Screen replay breaks when pages change	API first, captured calls next, browser last
Which steps are deterministic?	Totals and references must be exact	Reviewer locks steps to fixed rules
What needs human approval?	Payments and filings carry legal risk	All write actions gated by default
Whose credentials does it use?	Shared logins break accountability	Each user’s own login and permissions
What happens when a system changes?	Silent errors are worse than failures	Run stops, logs and rebuilds the path
Do model providers keep your data?	Payroll and KYC data are sensitive	Zero-data-retention with all three providers
What does each run cost?	Agent costs can exceed the labour saved	Cost shown per run; no list price
Who is the EU AI Act deployer?	Oversight duties follow the deployer	The customer

Map the workflow before the pilot

Akai learns from a recording, which means it learns whatever the recorded person does, including their workarounds. Documenting the process first, and deciding which version is correct, avoids automating a bad habit at scale. Our comparison of process mining and process mapping explains how to choose the right discovery method.

Pilot on the residual queue

Deel’s strongest results came from the cases its first-line automation could not clear, such as expense claims with missing receipts or unusual categories. Picking that queue for a back-office automation pilot gives a clear baseline, a measurable result and low risk, because a person was reviewing those cases anyway.

Plan for monitoring, not just deployment

Agents that run unattended need oversight after launch as well as before it. The recent run of incidents involving rogue agents, and the new tools designed to contain them such as Nvidia’s open agent safety platform, show why. Budget time for reviewing logs, handling exceptions and re-approving workflows when a process changes. Our guide to preventing automation project failure covers the operating model that keeps a back-office automation programme healthy after its first month.