AI in HR: SAP Agrees to Buy TechWolf for Skills Data, Workday Customers Use Outside Agents, and HackerRank Releases an AI Interviewer
Title: AI in HR: SAP Agrees to Buy TechWolf for Skills Data, Workday Customers Use Outside Agents, and HackerRank Releases an AI Interviewer
Date: 8 October 2026
Type: Paper
Author: SAASiQ (contact@saasiq.ai)
Word count: 2953 words
Reading time: 11 min
Published: 08-10-2026
SAP agreed on 6 October to buy TechWolf, a Belgian company that works out employees' skills from the work they do in business systems. SAP plans to use TechWolf's data to ground the HR agents in SuccessFactors. On 5 October Workday published a report built partly on hiring data from about 550 employers. The same day HackerRank made its AI interviewer, Chakra, generally available after about six months in beta. On 6 October Crypto Briefing summarised a report by The Information that Workday customers are using agents from Anthropic, Microsoft and others to pull data out of Workday. This paper sets out how each use works, what is known about cost, and where it breaks.
What happened this week
This is the second of SAASiQ's Thursday papers on how AI is used in practice, sector by sector. Finance came first on 1 October. This week is HR: hiring, skills, internal moves and the everyday work of managers.
SAP announced its agreement to buy TechWolf on Tuesday 6 October, as SAP's own news site, CIO, Techzine and The New Stack reported. Workday released its Global Workforce Report on Monday 5 October. HackerRank's release of Chakra was reported by TechCrunch on 5 October. The report on Workday's customers came from The Information and was summarised on 6 October by Crypto Briefing and Dealroom.
Two more items matter for the risk side. On 25 September the law firm Duane Morris reported that job applicants suing Workday had asked a federal court in California to certify their case as a class action. And on 1 October TechRadar reported a survey by Headway of 1,000 managers on how they use AI in reviews and feedback.
Use case one: working out who can do what
Most HR systems hold a skills list for each person, filled in by the employee and rarely updated. Workforce planning then rests on self-reported data that may be years old.
TechWolf takes a different route. According to SAP's announcement and the reports of 6 October, TechWolf builds what it calls a context graph for work. The graph has three layers: the tasks inside each role, the skills employees use, and outside labour market data. TechWolf infers skills from the tasks people perform in the business applications a company already runs. Its own product pages name Jira, Salesforce and ServiceNow among the sources. The inferred skills can then be checked by the employee and written back into the HR system's skill profiles.
SAP wants the graph as a grounding layer for Joule, the AI assistant across its applications. Manoj Swaminathan, president and chief product officer of SAP Autonomous Suite, said the graph would make token use more efficient and lower the cost of running workforce agents, according to the reports. SAP names skills-based hiring, workforce planning and role redesign as uses. SAP intends TechWolf to become the core of the SuccessFactors portfolio.
The terms were not disclosed. SAP expects the deal to close in the fourth quarter of 2026, subject to regulatory approval. TechWolf will keep its brand, its Ghent headquarters and its offices in London and New York. Its co-founder Andreas De Neve stays as chief executive, and TechWolf will go on serving customers that do not run SAP. TechWolf raised $42.75m in a Series B round led by Felix Capital in June 2024, and SAP took part in that round.
What the hiring data says about skills
Workday's report of 5 October drew on three sources. The first is a survey of 6,001 employees and business leaders, about 1,780 of them decision-makers. The second is skills data from job requisitions at about 550 employers using Workday Recruiting, from September 2025 to July 2026. The third is de-identified workforce data from Workday customers with at least 250 employees, matched year on year.
Demand for basic AI skills in job postings, such as simple prompting, peaked in January 2026 and then fell 25 per cent. Demand for hands-on skills, such as building AI tools, automating workflows and AI engineering, rose 51 per cent between September 2025 and July 2026.
On headcount, 40 per cent of leaders expect AI to help them get more from the staff they have. 28 per cent expect it to reduce headcount. The workforce data shows less movement inside organisations: internal moves fell at 57 per cent of employers compared with the year before, and promotion rates were roughly flat. Nearly four in ten employees reported a reorganisation or restructuring in the past year.
Workday sells HR software, so it has an interest in the subject. The skills and mobility figures come from its own customers, which are larger organisations.
Where skills inference breaks
The first weak point is coverage. A skills graph can only see work that passes through a connected system. An engineer who logs every ticket in Jira shows up well. A nurse, a site manager or a payroll clerk working in a system nobody connected shows up badly or not at all. If the graph then feeds hiring or redeployment, the people it cannot see lose out.
The second is consent and purpose. Inferring skills from work records means reading those records. Under UK GDPR an inferred skill about a named person is personal data, and employees have the right to know it is being made. A tool sold for career development can drift into use for redundancy selection. In SAASiQ's view, the purpose should be written down before the connectors are switched on, and any change of purpose treated as a new decision.
The third is drift. Roles change faster than skills taxonomies. If the graph is not kept current, it goes on matching people to jobs that no longer exist in that form. The practical check is to ask employees to confirm or correct their inferred skills, and to measure how often they correct them. A high correction rate means the inference is weak.
Use case two: agents that read HR data
HR suites now sell their own AI agents. Customers can also point general-purpose agents at the HR system through its APIs, the interfaces one system uses to call another. The Information's report, as summarised on 6 October, says Workday customers are doing more of the second.
The report rests on interviews with five consultants and partners who work with more than 2,000 Workday customers. Customers are using agents from Anthropic, Microsoft and others to pull data out of Workday, rather than buying or using the AI built into Workday. One consultancy with more than 400 Workday customers said about half had signed usage-based AI agreements. Only 10 to 20 per cent actively used Workday's AI, and none used Sana Enterprise, the AI product Workday launched in March. Software companies including NinjaOne and Druva are building their own agents on Workday data.
Workday's own figures point the other way. It says 5,500 of its roughly 11,500 customers use its AI agents, up 35 per cent on the previous quarter, without saying how many pay. Its AI products bring in close to $600m in annual recurring revenue. According to the report, Workday is offering some large customers free credits and a free year of Sana Enterprise. It has also waived API overage charges from May 2026 to February 2027. Workday's revenue growth slowed to 12.8 per cent in the quarter to July.
Workday's answer to outside agents is to register them. Its Agent System of Record is a list of every agent allowed to act in the system, Workday's own and anyone else's. Since an agreement announced in September 2025, agents built in Microsoft's Azure AI Foundry and Copilot Studio can be registered there. Each one gets a Microsoft Entra Agent ID, which gives it an identity of its own rather than a borrowed login.
The figures in the report come from consultants, and one consultancy's view is not a representative sample. They match the OvationCXM survey of US finance leaders that SAASiQ covered on 1 October, in which buyers wanted their own agents to reach their bank's data.
Where outside agents break
HR data is among the most sensitive data an organisation holds. HR systems hold salary, health, disciplinary and bank details in one place. An agent that can answer a manager's question about leave balances may, through the same connection, be able to read far more.
The control is scope. Each agent should have its own identity, its own permissions and its own log, separate from the person who set it up. Without them, an agent acting through a service account looks the same in the logs as every other integration, and an audit cannot say which tool read which record.
The second risk is cost. Usage-based AI agreements and API limits mean that an agent asking many small questions can cost more than a person running one report. Workday's waiver of overage charges runs to February 2027. In SAASiQ's view, any HR team running outside agents should find out now what those calls will cost once the waiver ends, and what its own contract says about API limits.
The third risk is the same one finance teams face. If the supported route is too slow or too expensive, staff build their own.
Use case three: screening and interviewing candidates
HackerRank, which sells coding assessments to more than 3,000 business customers, made Chakra generally available on 5 October. During about six months of testing it ran more than 500,000 interviews, according to the reports. Snowflake, Snorkel and Capgemini were among the companies that tested it.
Chakra changes the format of the technical test. The candidate gets a task set in a real-world code repository and works through it in an editor that includes an AI assistant. While the candidate works, Chakra asks follow-up questions based on what they are doing, such as why they chose one approach over another. It scores what HackerRank calls AI fluency alongside judgement and problem solving. HackerRank says Chakra scores candidates and people make the hiring decision.
LinkedIn announced a second version of its Hiring Assistant agent at its Talent Connect event, as HR Brew reported on 30 September. LinkedIn says the new version remembers a recruiter's past hiring patterns, such as location and seniority. It also shows signals of how receptive a candidate may be, based on platform activity, Open to Work settings and how often they answer InMail messages. Voice pre-screening is part of the update. Recruiters can switch individual steps on or off, override its recommendations and review what it has done.
CNBC reported on 15 September that some job seekers are refusing AI interviews and dropping employers from their search. A Greenhouse survey published in April found that 63 per cent of candidates had completed an AI-led interview, and HR Dive reported from the same work that 70 per cent said they were never told in advance that AI would assess them.
Where screening breaks
The legal test case is Mobley v. Workday in the Northern District of California. In May 2025 Judge Rita Lin allowed a nationwide age discrimination claim to proceed as a collective action for applicants aged 40 and over who used Workday's platform from 24 September 2020. Duane Morris reported on 25 September that the plaintiffs have now asked the court to certify a much wider class. They argue disparate impact under Title VII, the Americans with Disabilities Act and California's Fair Employment and Housing Act. The products named include HiredScore Spotlight, Fetch and Candidate Skills Match. Workday's position is that its tools do not make hiring decisions and that customers keep full control of their hiring. The court has not ruled on the new motion.
An earlier ruling in the case accepted that a software supplier could be treated as the employer's agent and so be directly liable for discrimination. If that holds, both the employer and the supplier carry the risk.
The UK has its own test. The Information Commissioner's Office published a report on automated decisions in recruitment on 31 March 2026, based on its work with more than 30 employers. Many described their tools as decision support. The ICO's view was that a number of them were in fact making solely automated decisions. Privacy notices were usually too general for candidates to understand how their data was used. The ICO wrote to 16 organisations with recommendations. Its draft guidance says a person who rubber-stamps the tool's output does not count as human review: the reviewer has to have the authority and the competence to change the outcome.
The rules behind that changed on 5 February 2026, when the automated decision provisions of the Data (Use and Access) Act 2025 came into force. The old near-ban on significant solely automated decisions became a set of safeguards: tell the person, let them contest it, and offer human review. The tightest limits now apply mainly where special category data is used. The ICO's consultation on its draft guidance closed on 29 May, and its forward plan lists the final version for winter 2026.
Elsewhere, New York City already requires an annual independent bias audit of automated hiring tools, a published summary and ten business days' notice to candidates. Colorado replaced its 2024 AI Act in May 2026 with a narrower law that requires notice, an explanation of adverse decisions within 30 days and human review. California's notice rules for automated decision tools apply from 1 January 2027. Under the EU AI Act, recruitment tools are high-risk under Annex III, and those obligations apply from 2 December 2027.
Use case four: managers writing reviews
Headway's survey of 1,000 managers, reported by TechRadar on 1 October, found that 33 per cent had used AI to write feedback or prepare a performance review. That is down from 41 per cent in a similar survey a year earlier. 31 per cent had used AI to help resolve conflicts in their team.
Where this happens outside the HR system, a manager pastes notes into a chatbot and pastes the result into the review form. The organisation's records then show a review the manager signed, with no sign of what the tool added or what data it was given.
Reviews feed pay, promotion and, at the hard end, dismissal. A review drafted by a tool that never saw the employee's work is weak evidence if it is ever challenged. In SAASiQ's view, the simplest control is a plain rule: managers may use an approved tool to draft, they must not paste personal data into an unapproved one, and the final text is theirs.
What it costs and who pays
SAP did not disclose what it is paying for TechWolf, and the reports gave no price for Chakra or the new Hiring Assistant.
The cost structure is clearer. Suites are moving AI onto usage-based terms: credits, metered agent actions and API charges. The Information's report shows customers signing those agreements and then not using them, and Workday answering with free credits and waived overages. For a buyer, that means two separate costs to track. One is the AI bought from the suite supplier. The other is the cost of outside agents reaching the same data, which shows up as API use on one bill and model use on another.
SAP's case for TechWolf is partly a cost case: better grounding means fewer tokens per answer, which lowers the running cost of each agent. That claim is SAP's own. No figures were published with it.
The costs that do not appear on any invoice are the controls: bias audits, human reviewers with real authority, candidate notices, logs and the people who read them. In SAASiQ's view these belong in the business case from the start, because the legal cases above are about what happens when they are missing.
What an HR team can do on Monday
List every AI feature that touches people decisions. Include features that arrived in a quarterly update of the HR suite, add-ons in the applicant tracking system, assessment tools such as coding tests and video interviews, and any outside agent with access to HR data. For each one, record what it reads, what it produces, and who acts on the result.
Apply the ICO's test to each one. Does a person with the authority and knowledge to change the outcome look at every case before it takes effect? If not, treat it as a solely automated decision and make sure candidates or employees are told, can contest it and can get human review.
Measure the overrides. For any screening or ranking tool, record how often recruiters reject its recommendations, and check the pass rates by age, sex and other protected characteristics where the data allows.
Register outside agents. Any agent reading the HR system should have its own identity, the narrowest permissions it needs, and logs that show what it read. Whichever HR suite is in use, ask the supplier how it identifies and limits agents that it did not build.
Read the contract. Check how AI use is metered, what the API limits are, and when any free credits or waivers end. Ask the supplier for its most recent bias audit of each screening feature, and what it says about liability if a claim is brought.
Tell candidates. State in the job advert or application process where AI is used and how to ask for a person instead. New York City and Colorado already require it.
What happens next
Workday holds its Rising conference in Las Vegas from 12 to 15 October. SAP expects the TechWolf deal to close in the fourth quarter of 2026. The court in Mobley v. Workday has yet to rule on class certification, and the ICO's final guidance on automated decisions is listed for winter 2026.
SAASiQ - Intelligent Solutions for SaaS ©
