Skip to main content

Building a Bob Data Audit & Reporting Workflow with Bob MCP in Claude

  • Sep 25, 2026 7:24 PM UTC
  • 14 replies
  • 695 views
  • Translate All

Issabella Luna

Hi All,

I’m Issabella (Issa), a People Ops Analyst and I am responsible for maintaining and managing the HRIS. Nice to virtually meet you! Its been about 2 months since we’ve enabled Bob MCP access in Claude company-wide, and the results have been astounding! We have been seeing great use out of the business and truly have seen the tool take our efficiencies and processes to the next level.

As the person that manages the HRIS I am responsible for owning data integrity in the system, which before MCP was a very manual and tedious task that took place in spreadsheets and reports. This was one vein of work that needed to be automated and the use case was the perfect candidate for AI. Using the MCP in Claude, I built a data audit workflow and “dashboard” that would allow me to track data points being entered that were not consistent with our system architecture rules.

After some testing, it is running weekly and sending out messages to the data entry users to correct anything that is out of place. It does still need some work and the I am adding flags and rules to the logic as things pop up. If I had to guess though, I am saving roughly an average of 4 hours per week with where its at today. 

I thought I would share my methodology breakdown and an example dashboard here (using dummy data) in case anyone wanted to try to build something out that is similar! Reach out with any questions, thanks for reading!

14 replies

Hi All,

I’m Issabella (Issa), a People Ops Analyst and I am responsible for maintaining and managing the HRIS. Nice to virtually meet you! Its been about 2 months since we’ve enabled Bob MCP access in Claude company-wide, and the results have been astounding! We have been seeing great use out of the business and truly have seen the tool take our efficiencies and processes to the next level.

As the person that manages the HRIS I am responsible for owning data integrity in the system, which before MCP was a very manual and tedious task that took place in spreadsheets and reports. This was one vein of work that needed to be automated and the use case was the perfect candidate for AI. Using the MCP in Claude, I built a data audit workflow and “dashboard” that would allow me to track data points being entered that were not consistent with our system architecture rules.

After some testing, it is running weekly and sending out messages to the data entry users to correct anything that is out of place. It does still need some work and the I am adding flags and rules to the logic as things pop up. If I had to guess though, I am saving roughly an average of 4 hours per week with where its at today. 

I thought I would share my methodology breakdown and an example dashboard here (using dummy data) in case anyone wanted to try to build something out that is similar! Reach out with any questions, thanks for reading!

This is fantastic. 

Translate

Harshil Boparai
Bobber

Wow, ​@Issabella Luna , this is brilliant! Taking a manual data audit and turning it into a weekly workflow that flags issues for the right people to fix is such a practical use of Bob MCP. Saving around four hours a week is impressive, too. Thank you for sharing your approach and example dashboard so others can learn from it! 👏

Translate

Love what you build Issabella! I’ll definitely give it a try myself and see if I can get it to work as well - Might reach out if I walk into problems - Thank you for sharing!

Translate

Antonia Steele
Bobber

Thanks so much for sharing this ​@Issabella Luna It’s fantastic to see the impact MCP is already having and such a great real-life example of how it can be used to automate a previously manual process.

Translate

Paul Rousell

This is great and inspiring! Thanks so much for sharing ​@Issabella Luna. 

Translate

Wow, this was great! 
Super cool use case. 

Did anyone try to recreate it for themselves yet? 

Translate

​@Issabella Luna This is fantastic. I have a question. Do you create a specific Service Users in Hibob to connect your MCP with Claude? Our security is extremely conservative, and I have not gotten approval to connect HiBob MCP with Claude due to confidentiality and security concerns. 

How do you go about ensuring Claude pull only specific info through MCP for specific purpose? 

 

Translate

Issabella Luna

​@LRodriguez Great question! We have connected MCP to Claude via OAuth. Its all permission dependent, at the user-level, based on the have permissions configured for employees in system. So if you feel confident about your permissions set in HiBob, there is no risk of confidential data being exposed. Additionally, once its connected in Claude you can set restrictions to the tools that are available to use with the MCP (read and write paths). So if you did not want to allow employees to submit a time off request from the MCP connection with Claude, you can toggle that specific tool path to “Block”. Hope this helps!

Translate

This is very helpful! 
I mirrored it to create a Probationary Review Tracker so we can see who’s pending their Probationary Reviews.

Translate

Issabella Luna

​@Nelly Kapen Love this idea!

Translate

Thank you for getting the ball rolling! I truly appreciate the work you did 

Translate

Hey Issabella - thanks for sharing this, really enjoyed reading it! The single tracker every step reads from, writing the exceptions down with the reasoning behind them, and tracking new vs still open are all things we landed on too.

 

One thing I’d gently push on, and it’s very much a call for your team - is how hands-on it seems. Right  now it looks like nothing gets finished without you, since every nudge is a draft you read, sort and send yourself. We’ve ended up running ours the other way round: the automation does the work and we check it afterwards. A person only signs off upfront where a mistake really hurts.

 

How we’re doing it ourselves at Easygo
Ours is a bit of a bigger, custom build, so not all of it will translate - but the principles might. We’ve built a team of AI agents, each with one job, running on a schedule by themselves and posting as themselves in our HR Ops Slack channel. For HiBob data:

  • Our data-check agent reads every HiBob record on a loop and checks it against a fixed set of rules plus our own house rules. Each new issue gets its own Slack thread with what’s wrong, the likely cause and the suggested fix. Nobody kicks it off and nobody approves the alerts.
  • If the fix is a HiBob change, it posts a ready-to-go card. One ✅ from an approver and it makes the change, then reads it back to confirm it landed. It can only ever change two fields - anything else goes to a person.
  • Exceptions teach themselves. If someone replies “that’s a genuine rehire, it’s fine”, the agent suggests it as a standing rule. Once it’s ticked, it never flags that case again - no one has to go back and edit a playbook.
  • A separate auditor agent checks every change any agent makes within about 15 minutes against what’s actually in HiBob and Jira, and marks the Slack card once it matches. If an agent gets something wrong, it shows up on its own.
  • Each agent has a written rules file and a short list of the only things it’s allowed to change. Every decision I make goes into a log the agents work from, and every change to an agent gets built and tested on a copy before it goes live.
  • Approvals before the fact are kept for pay, visa and immigration, ER matters, and changes to someone’s record. Everything else runs and gets checked afterwards. As an agent proves itself, we turn the asking down - we’ve just set one to stop asking for a tick once 10 in a row are right, and to start asking again after a single miss.


What that could look like for you

  • Let the low-risk nudges send themselves - a missing field, an out-of-date address. A wrong nudge costs one confused colleague, whereas approving every single one costs you time every week.
  • Send them from a bot rather than your own account - people know it’s automated and nothing goes out in your name.
  • Keep your sign-off for pay, visa and anything sensitive. Leaving salary outliers out of the nudges altogether is spot on.
  • Swap the upfront approval for checking afterwards - next week’s run confirms the record actually got fixed, and you read one weekly summary instead of every message.
  • If you’d rather ease into it, keep approving for a few weeks and count how many drafts go out unchanged. Once a category is consistently right, let it run.
     

A few other things we learned the hard way

  • Only pull pay if a rule actually needs it - otherwise the dated pull files leave salary sitting on disk every week.
  • Make a missed run loud - a weekly job that quietly stops looks exactly like a clean week. Have each run say “clean”, “found issues” or “couldn’t check, and here’s why”, and only get pinged when something’s late or wrong. Worth checking the schedule still fires when your laptop’s asleep too!
  • Test any rule change on a handful of fake bad records (a rehire, a missing manager) so a tweak can’t quietly stop catching things.
  • Treat free text in HiBob as data, not instructions - if a field ever reads like a command, have Claude flag it rather than act on it.

Apologies if I misinterpreted anything, but hopefully some of the above may be useful!

Translate

Issabella Luna

​@Jack Amerena Thanks for reading, Jack. And thank you for all of your great suggestions! I would love to get to a point where we could build an agent or 2 to take some of the low-risk load. This is definitely a work in progress project and will always keep evolving, but I think the natural next step is having it automatically sending the nudges for me. Would you be open to connecting on how you built and incorporated agents into your workflow? My linkedin is here. Thanks and have a great weekend!

Translate

​@Issabella Luna - that is an awesome piece of work you and your team have created. I’m with you in asking ​@Jack Amerena to show us his AI agent building skills and how he went about it! 

Translate