Grab
Grab's procedure-driven AI agents cut fraud and account-takeover investigation time
Grab built an AI agent framework that follows written standard operating procedures, and used it to automate account takeover investigations and fraud checks in risk operations. Grab reports a drop in investigation time from 23 minutes to 3 and 87% of cases automated.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
87%
Share of account takeover investigation cases automated by the ATI bot1
45%
Reduction in average handling time for fraud investigation1
over 300
Man-hours saved each month in fraud investigation1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Investigating an account takeover, where someone else gains control of a customer's account, was slow manual work. Analysts pulled data from several systems, ran many database queries and made decisions under pressure, which led to delays, mistakes and a need for specialist skills.
What they built
Grab built a framework in which an AI agent works through a written standard operating procedure (SOP). The first use was the Account Takeover Investigations (ATI) bot, and the same framework was applied to fraud investigation. Staff review results in tools such as Slack.
How they did it
The SOP is written in plain language as a tree of steps and decisions, and non-technical staff can edit it in a visual editor. A planner module walks the tree step by step. A worker agent runs the database queries and system calls each step requires and stores the results. At the end the system writes a summary and a decision. Grab limits the tools the model can use and compresses context to reduce invented steps.
Results
Grab reports that the ATI bot cut investigation time from 23 minutes to just 3 and automated 87% of cases. Fraud investigation average handling time fell by 45%, saving over 300 man-hours a month. A second Grab post repeats the 87% figure and gives the starting time as 22 minutes rather than 23.
What to copy
Start with a process that already has a written procedure and clear decisions. Give process owners control of the procedure, and keep a record of each run.
Lessons for your project
Encode the human procedure as a step-by-step tree, with decision points and data calls. Grab constrains the model with the written SOP instead of letting it improvise.
Let non-technical process owners write and edit the procedure. Grab gives them a visual editor so the people who know the process control it.
Keep humans in the loop where needed. The framework can pause and resume so a person can step in.
Log every run so decisions can be explained and reviewed.
Caveats
- All figures are Grab's own, from two engineering blog posts. No outside party confirms them.
- The first post gives 23 minutes before the bot and the second gives 22 minutes, both reduced to 3 minutes per ticket. Grab does not explain the difference, and only the first post's figures are listed as metrics.
- Grab calls the over 99.8% accuracy and the 0 false rate results of initial evaluations. They are not listed as metrics here. The posts do not describe how they were measured.
- The posts do not name the AI models, and give no launch date. They describe deployed bots, and the deployed date shown is the date of the first post.
- The posts do not say how many cases or analysts the figures cover, or what the build cost.
Sources
- Introducing the SOP-driven LLM agent frameworks
Grab Engineering Blog · 2025-04-25 · Company blog · engineering.grab.com
- Streamlining RiskOps with the SOP agent framework
Grab Engineering Blog · 2025-05-08 · Company blog · engineering.grab.com
Summarized in our own words from the sources above. Spotted an error? See corrections.