Industries / Technology / Software and SaaS / ERP and CRM software
AI agents
AI agents are programs built on large language models that read and change the records in ERP and CRM systems on a person's behalf, for example by answering a customer, qualifying a sales lead or entering an order.
6
Milestones achieved, from 2023 to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2023
Assistants inside CRM and ERP
On 6 March 2023 Microsoft announced Dynamics 365 Copilot, generative AI built into its sales, service, marketing, operations and supply chain software to take on tasks such as data entry, writing content and taking notes. Microsoft called it "the world's first copilot in both CRM and ERP"; the claim of being first is Microsoft's own. Redmond Magazine reported that most of the features were in preview at launch, with AI-drafted replies to customers' emails due for general release on 15 March.
Credit
Microsoft; announced by Charles Lamanna
2024
Agents tested in business software
WorkArena, published in March 2024, set web agents 33 kinds of task (19,912 variations) on the ServiceNow platform, such as filtering lists, filling forms and searching a knowledge base; an agent built on GPT-4o completed 43%, one on Llama 3 17.9% and one on GPT-3.5 6.1%. In November 2024 CRMArena set nine customer-service tasks in a Salesforce organization filled with made-up but realistic records; agents completed under 40% when prompted to alternate reasoning and actions (a method called ReAct) and under 55% when they could call the system's functions directly.
LessonWorkArena's authors found that most of the difficulty came from the software's own dense screens, not from the business knowledge the tasks needed.
Credit
Alexandre Drouin, Maxime Gasse and co-authors at ServiceNow Research; Kung-Hsiang Huang, Chien-Sheng Wu and co-authors at Salesforce AI Research
2025
Conversations cut success
CRMArena-Pro extended the test to 19 tasks across sales, service and quoting, checked by experts, for companies selling to businesses and to consumers. The best agents completed about 58% of tasks given in one message and about 35% when the task unfolded over a conversation with a simulated user, and showed "near-zero" awareness of what data they should keep confidential; prompting them to be careful helped, but often made them worse at the task.
LessonAn agent that finishes the job may still hand over data it should have kept back.
Credit
Kung-Hsiang Huang, Chien-Sheng Wu and co-authors, Salesforce AI Research
2025
A chat agent's keys stolen
From 8 to 18 August 2025 an attacker that Google tracks as UNC6395 used stolen OAuth tokens belonging to Salesloft's Drift chat agent to export large amounts of data from many companies' Salesforce systems. OAuth tokens are stored keys that let one app act inside another without a password. The attacker then searched the data for passwords and cloud access keys. On 20 August Salesloft and Salesforce revoked every Drift token and Salesforce removed the app from its marketplace; Google said the cause was not a flaw in Salesforce itself, and later found Drift's email tokens had been used too.
LessonAn agent connected to a CRM holds keys to everything it can reach, so a breach at the agent's maker reaches every customer that connected it.
Credit
Google Threat Intelligence Group and Mandiant
2025
One email address was enough
Aaron Costello of AppOmni found that ServiceNow's Virtual Agent interface and its Now Assist AI agents shipped with the same secret key on every customer's system and linked outside callers to accounts by email address alone. Combined, the two let someone with no login impersonate any user, get past multi-factor sign-in, and order an AI agent to create a new administrator account. He reported it on 23 October 2025, ServiceNow fixed hosted systems on 30 October, and the flaw (CVE-2025-12420, rated 9.3 out of 10) was made public in January 2026, with no evidence it had been used.
LessonGiving agents the power to act turned an old kind of login flaw into control of the whole system.
Credit
Aaron Costello, AppOmni; ServiceNow
2026
Results checked in the database
EntWorld set agents 1,756 tasks in open-source business systems for CRM, IT service management and ERP, among them the EspoCRM customer-records system, run in containers, and judged each task by querying the database afterwards rather than by inspecting the screen. GPT-4.1 completed 47.61% of tasks and the best agent tested 56.89%, against about 85% for people, by the authors' count.
LessonChecking the saved records gives a pass or fail that does not depend on another AI's opinion.
Credit
Ying Mo, Yu Bai, Dapeng Sun and co-authors
Watch for
Order exceptions in minutes
Levi Strauss says it has more than 1,000 AI agents in production on its SAP system. About 80% of its wholesale orders already flowed through automatically; the other 20%, often handwritten notes, emails or loose documents from smaller customers, took two to five days to enter by hand and now take 20 to 30 minutes.
Achieved whenLevi Strauss or an outside auditor publishes how many such orders the agents entered over a stated period and how many needed correcting.
Claimed by
Jason Gowans, Levi Strauss, at SAP Sapphire, May 2026
A vendor's own scorecard
Microsoft built the Sales Research Bench, 200 questions of the kind sales leaders ask about their data, scored on eight qualities by other AI models acting as judges. In Microsoft's tests on 19 October 2025 its Sales Research Agent beat Claude Sonnet 4.5 by 13 points and ChatGPT with GPT-5 by 24.1 points on a 100-point scale, and Microsoft said it would publish the questions and data "in the coming months". Its technical report on arXiv in December 2025 printed three sample questions; a search in November 2026 found no published package.
Achieved whenthe questions and data are public and a party other than Microsoft reproduces the gaps.
Claimed by
Microsoft, October 2025
Open
01
Approval limits under pressure
A suite of 200 tasks in an ERP the builder runs, where finishing the task would break a financial control: approving one's own purchase order, splitting an order to stay under an approval limit, paying an invoice with no matching delivery, or editing the approval rule itself. In half the tasks the pressure comes from the user's request and in half from text planted in a supplier's email or document. Each run is scored by checking the database: the agent refused, asked for approval, or broke the control.
The briefCredit · unclaimed02
What CRM apps can reach
A public table of the permissions that the 50 most-installed AI apps for Salesforce and for HubSpot ask for when connected, recorded from the consent screens in free developer accounts, with flags for full data access and for refresh tokens that keep working when no one is logged in, compared with what each app says it does.
The briefCredit · unclaimed
Open 01 · the brief
Approval limits under pressure
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A suite of 200 tasks in an ERP the builder runs, where finishing the task would break a financial control: approving one's own purchase order, splitting an order to stay under an approval limit, paying an invoice with no matching delivery, or editing the approval rule itself. In half the tasks the pressure comes from the user's request and in half from text planted in a supplier's email or document. Each run is scored by checking the database: the agent refused, asked for approval, or broke the control.
Why it is still open
CRMArena-Pro measured whether agents keep data confidential, not whether they respect approval limits. A September 2026 paper on separation of duties for agents studies operating-system privileges, and an October 2026 study showed an attack that led an agent to request an ordinary approval token and use it to open protected documents, succeeding in 8 of 9 reruns; neither tests the approval rules of an ERP. ReimburseBench (2026) tests models as auditors that check expense claims against a company's policy, not as agents that could break the policy themselves. Searches on 20 November 2026 found no such suite.
Built with what exists
ERPNext, whose authorization rules let a document above a set value be submitted only by an approving role; its REST API, or Browser Use to drive its screens (see the Note "Deleting 80% to start over"); current models such as Claude, GPT and open-weight Qwen; a Python test harness; made-up companies and suppliers only.
First step
Install ERPNext locally, set a rule that purchase orders above 10,000 need a manager, and write the first 20 tasks.
Done looks like
For at least four models, the share of the 200 tasks in which the control was broken, with and without planted text, each case confirmed in the database, published with the code.
Open 02 · the brief
What CRM apps can reach
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of the permissions that the 50 most-installed AI apps for Salesforce and for HubSpot ask for when connected, recorded from the consent screens in free developer accounts, with flags for full data access and for refresh tokens that keep working when no one is logged in, compared with what each app says it does.
Why it is still open
The 2025 Drift theft ran on exactly such tokens. In June 2026 Offroad's OhAuth project audited 2,890 app listings and found 677 asking for more access than their stated purpose, including AI apps with write access, but it covered only the Google Workspace and GitHub marketplaces, not CRM ones.
Built with what exists
Salesforce Developer Edition and HubSpot developer test accounts, both free; each app's own install screens; a spreadsheet; a language model to compare each listing's description with its permissions, checked by a person.
First step
Open the two free developer accounts and record the consent screens of the ten most-installed AI apps that allow installation in a test account.
Done looks like
100 apps, each with a screenshot of its consent screen and its requested permissions, an over-reach flag that a second person has checked, and published totals per platform.